From 86d989ed0862c7700d8fd25aa8d2e954aa9618c9 Mon Sep 17 00:00:00 2001 From: nipeng Date: Wed, 16 Oct 2019 09:34:32 +0800 Subject: [PATCH 01/23] =?UTF-8?q?=E4=BF=AE=E6=94=B9=E9=85=8D=E7=BD=AE?= =?UTF-8?q?=E6=96=87=E4=BB=B6=EF=BC=8C=E8=B7=91=E8=87=AA=E6=9C=89=E6=95=B0?= =?UTF-8?q?=E6=8D=AE?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- egs/aishell/asr1/conf/fbank.conf | 2 +- egs/aishell/asr1/conf/pitch.conf | 2 +- egs/aishell/asr1/infer.sh | 255 ++++++++++++++++++ egs/aishell/asr1/local/aishell_data_prep.sh | 2 +- .../asr1/local/aishell_data_prep_infer.sh | 66 +++++ egs/aishell/asr1/run.sh | 25 +- .../tts1/conf/tuning/train_fastspeech.v1.yaml | 2 +- .../tuning/train_pytorch_transformer.v1.yaml | 2 +- egs/ljspeech/tts1/local/text/cleaners.py | 4 +- egs/ljspeech/tts1/run.sh | 24 +- tools/Makefile | 10 +- 11 files changed, 361 insertions(+), 33 deletions(-) create mode 100755 egs/aishell/asr1/infer.sh create mode 100755 egs/aishell/asr1/local/aishell_data_prep_infer.sh diff --git a/egs/aishell/asr1/conf/fbank.conf b/egs/aishell/asr1/conf/fbank.conf index 82ac7bd0dbc..1ad20614eef 100644 --- a/egs/aishell/asr1/conf/fbank.conf +++ b/egs/aishell/asr1/conf/fbank.conf @@ -1,2 +1,2 @@ ---sample-frequency=16000 +--sample-frequency=8000 --num-mel-bins=80 diff --git a/egs/aishell/asr1/conf/pitch.conf b/egs/aishell/asr1/conf/pitch.conf index e959a19d5b8..926bcfca92a 100644 --- a/egs/aishell/asr1/conf/pitch.conf +++ b/egs/aishell/asr1/conf/pitch.conf @@ -1 +1 @@ ---sample-frequency=16000 +--sample-frequency=8000 diff --git a/egs/aishell/asr1/infer.sh b/egs/aishell/asr1/infer.sh new file mode 100755 index 00000000000..b9cbb023099 --- /dev/null +++ b/egs/aishell/asr1/infer.sh @@ -0,0 +1,255 @@ +#!/bin/bash + +# Copyright 2017 Johns Hopkins University (Shinji Watanabe) +# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) + +. ./path.sh || exit 1; +. ./cmd.sh || exit 1; + +# general configuration +backend=pytorch +stage=0 # start from 0 if you need to start from data preparation +stop_stage=100 +ngpu=0 # number of gpus ("0" uses cpu, otherwise use gpu) +debugmode=1 +dumpdir=dump # directory to dump full features +N=0 # number of minibatches to be used (mainly for debugging). "0" uses all minibatches. +verbose=0 # verbose option +resume= # Resume the training from snapshot + +# feature configuration +do_delta=false + +train_config=conf/train.yaml +lm_config=conf/lm.yaml +decode_config=conf/decode.yaml + +# rnnlm related +lm_resume= # specify a snapshot file to resume LM training +lmtag= # tag for managing LMs + +# decoding parameter +recog_model=model.acc.best # set a model to be used for decoding: 'model.acc.best' or 'model.loss.best' +n_average=10 + +# data +#data=/export/a05/xna/data +#data_url=www.openslr.org/resources/33 + +data=/data/nipeng/2019-10-07-aishell + +# exp tag +tag="" # tag for managing experiments. + +. utils/parse_options.sh || exit 1; + +# Set bash to 'debug' mode, it will exit on : +# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands', +set -e +set -u +set -o pipefail + +train_set=train_sp +train_dev=dev +#recog_set="dev test" +recog_set="infer" + +# if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then +# echo "stage -1: Data Download" +# local/download_and_untar.sh ${data} ${data_url} data_aishell +# local/download_and_untar.sh ${data} ${data_url} resource_aishell +# fi + +if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then + ### Task dependent. You have to make data the following preparation part by yourself. + ### But you can utilize Kaldi recipes in most cases + echo "stage 0: Data preparation" + local/aishell_data_prep_infer.sh ${data}/data_aishell/wav ${data}/data_aishell/transcript + # remove space in text + for x in infer; do + cp data/${x}/text data/${x}/text.org + paste -d " " <(cut -f 1 -d" " data/${x}/text.org) <(cut -f 2- -d" " data/${x}/text.org | tr -d " ") \ + > data/${x}/text + rm data/${x}/text.org + done +fi + +feat_tr_dir=${dumpdir}/${train_set}/delta${do_delta}; mkdir -p ${feat_tr_dir} +feat_dt_dir=${dumpdir}/${train_dev}/delta${do_delta}; mkdir -p ${feat_dt_dir} +if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then + ### Task dependent. You have to design training and dev sets by yourself. + ### But you can utilize Kaldi recipes in most cases + echo "stage 1: Feature Generation" + fbankdir=fbank + # Generate the fbank features; by default 80-dimensional fbanks with pitch on each frame + # steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 32 --write_utt2num_frames true \ + # data/train exp/make_fbank/train ${fbankdir} + # utils/fix_data_dir.sh data/train + # steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 10 --write_utt2num_frames true \ + # data/dev exp/make_fbank/dev ${fbankdir} + # utils/fix_data_dir.sh data/dev + steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 10 --write_utt2num_frames true \ + data/infer exp/make_fbank/infer ${fbankdir} + utils/fix_data_dir.sh data/infer + + # # speed-perturbed + # utils/perturb_data_dir_speed.sh 0.9 data/train data/temp1 + # utils/perturb_data_dir_speed.sh 1.0 data/train data/temp2 + # utils/perturb_data_dir_speed.sh 1.1 data/train data/temp3 + # utils/combine_data.sh --extra-files utt2uniq data/${train_set} data/temp1 data/temp2 data/temp3 + # rm -r data/temp1 data/temp2 data/temp3 + # steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 32 --write_utt2num_frames true \ + # data/${train_set} exp/make_fbank/${train_set} ${fbankdir} + # utils/fix_data_dir.sh data/${train_set} + + # # compute global CMVN + # compute-cmvn-stats scp:data/${train_set}/feats.scp data/${train_set}/cmvn.ark + + # # dump features for training + # split_dir=$(echo $PWD | awk -F "/" '{print $NF "/" $(NF-1)}') + # if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d ${feat_tr_dir}/storage ]; then + # utils/create_split_dir.pl \ + # /export/a{11,12,13,14}/${USER}/espnet-data/egs/${split_dir}/dump/${train_set}/delta${do_delta}/storage \ + # ${feat_tr_dir}/storage + # fi + # if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d ${feat_dt_dir}/storage ]; then + # utils/create_split_dir.pl \ + # /export/a{11,12,13,14}/${USER}/espnet-data/egs/${split_dir}/dump/${train_dev}/delta${do_delta}/storage \ + # ${feat_dt_dir}/storage + # fi + # dump.sh --cmd "$train_cmd" --nj 32 --do_delta ${do_delta} \ + # data/${train_set}/feats.scp data/${train_set}/cmvn.ark exp/dump_feats/train ${feat_tr_dir} + for rtask in ${recog_set}; do + feat_recog_dir=${dumpdir}/${rtask}/delta${do_delta}; mkdir -p ${feat_recog_dir} + dump.sh --cmd "$train_cmd" --nj 10 --do_delta ${do_delta} \ + data/${rtask}/feats.scp data/${train_set}/cmvn.ark exp/dump_feats/recog/${rtask} \ + ${feat_recog_dir} + done +fi + +dict=data/lang_1char/${train_set}_units.txt +echo "dictionary: ${dict}" +if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then + ### Task dependent. You have to check non-linguistic symbols used in the corpus. + echo "stage 2: Dictionary and Json Data Preparation" + mkdir -p data/lang_1char/ + + # echo "make a dictionary" + # echo " 1" > ${dict} # must be 1, 0 will be used for "blank" in CTC + # text2token.py -s 1 -n 1 data/${train_set}/text | cut -f 2- -d" " | tr " " "\n" \ + # | sort | uniq | grep -v -e '^\s*$' | awk '{print $0 " " NR+1}' >> ${dict} + # wc -l ${dict} + + # echo "make json files" + # data2json.sh --feat ${feat_tr_dir}/feats.scp \ + # data/${train_set} ${dict} > ${feat_tr_dir}/data.json + for rtask in ${recog_set}; do + feat_recog_dir=${dumpdir}/${rtask}/delta${do_delta} + data2json.sh --feat ${feat_recog_dir}/feats.scp \ + data/${rtask} ${dict} > ${feat_recog_dir}/data.json + done +fi + +# you can skip this and remove --rnnlm option in the recognition (stage 5) +if [ -z ${lmtag} ]; then + lmtag=$(basename ${lm_config%.*}) +fi +lmexpname=train_rnnlm_${backend}_${lmtag} +lmexpdir=exp/${lmexpname} +# mkdir -p ${lmexpdir} + +# if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then +# echo "stage 3: LM Preparation" +# lmdatadir=data/local/lm_train +# mkdir -p ${lmdatadir} +# text2token.py -s 1 -n 1 data/train/text | cut -f 2- -d" " \ +# > ${lmdatadir}/train.txt +# text2token.py -s 1 -n 1 data/${train_dev}/text | cut -f 2- -d" " \ +# > ${lmdatadir}/valid.txt + +# ${cuda_cmd} --gpu ${ngpu} ${lmexpdir}/train.log \ +# lm_train.py \ +# --config ${lm_config} \ +# --ngpu ${ngpu} \ +# --backend ${backend} \ +# --verbose 1 \ +# --outdir ${lmexpdir} \ +# --tensorboard-dir tensorboard/${lmexpname} \ +# --train-label ${lmdatadir}/train.txt \ +# --valid-label ${lmdatadir}/valid.txt \ +# --resume ${lm_resume} \ +# --dict ${dict} +# fi + +if [ -z ${tag} ]; then + expname=${train_set}_${backend}_$(basename ${train_config%.*}) + if ${do_delta}; then + expname=${expname}_delta + fi +else + expname=${train_set}_${backend}_${tag} +fi +expdir=exp/${expname} +# mkdir -p ${expdir} + +# if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then +# echo "stage 4: Network Training" +# ${cuda_cmd} --gpu ${ngpu} ${expdir}/train.log \ +# asr_train.py \ +# --config ${train_config} \ +# --ngpu ${ngpu} \ +# --backend ${backend} \ +# --outdir ${expdir}/results \ +# --tensorboard-dir tensorboard/${expname} \ +# --debugmode ${debugmode} \ +# --dict ${dict} \ +# --debugdir ${expdir} \ +# --minibatches ${N} \ +# --verbose ${verbose} \ +# --resume ${resume} \ +# --train-json ${feat_tr_dir}/data.json \ +# --valid-json ${feat_dt_dir}/data.json +# fi + +if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then + echo "stage 5: Decoding" + nj=40 + if [[ $(get_yaml.py ${train_config} model-module) = *transformer* ]]; then + recog_model=model.last${n_average}.avg.best + # average_checkpoints.py --backend ${backend} \ + # --snapshots ${expdir}/results/snapshot.ep.* \ + # --out ${expdir}/results/${recog_model} \ + # --num ${n_average} + fi + pids=() # initialize pids + for rtask in ${recog_set}; do + ( + decode_dir=decode_${rtask}_$(basename ${decode_config%.*})_${lmtag} + feat_recog_dir=${dumpdir}/${rtask}/delta${do_delta} + + # split data + splitjson.py --parts ${nj} ${feat_recog_dir}/data.json + + #### use CPU for decoding + ngpu=0 + + ${decode_cmd} JOB=1:${nj} ${expdir}/${decode_dir}/log/decode.JOB.log \ + asr_recog.py \ + --config ${decode_config} \ + --ngpu ${ngpu} \ + --backend ${backend} \ + --batchsize 0 \ + --recog-json ${feat_recog_dir}/split${nj}utt/data.JOB.json \ + --result-label ${expdir}/${decode_dir}/data.JOB.json \ + --model ${expdir}/results/${recog_model} \ + --rnnlm ${lmexpdir}/rnnlm.model.best + + score_sclite.sh ${expdir}/${decode_dir} ${dict} + + ) & + pids+=($!) # store background pids + done + i=0; for pid in "${pids[@]}"; do wait ${pid} || ((++i)); done + [ ${i} -gt 0 ] && echo "$0: ${i} background jobs are failed." && false + echo "Finished" +fi diff --git a/egs/aishell/asr1/local/aishell_data_prep.sh b/egs/aishell/asr1/local/aishell_data_prep.sh index 4747e4f4d82..700afa3dee9 100755 --- a/egs/aishell/asr1/local/aishell_data_prep.sh +++ b/egs/aishell/asr1/local/aishell_data_prep.sh @@ -12,7 +12,7 @@ if [ $# != 2 ]; then fi aishell_audio_dir=$1 -aishell_text=$2/aishell_transcript_v0.8.txt +aishell_text=$2/aishell_transcript_v0.8.txt.0604.relabeled.0923.v2 train_dir=data/local/train dev_dir=data/local/dev diff --git a/egs/aishell/asr1/local/aishell_data_prep_infer.sh b/egs/aishell/asr1/local/aishell_data_prep_infer.sh new file mode 100755 index 00000000000..f4b75d102a7 --- /dev/null +++ b/egs/aishell/asr1/local/aishell_data_prep_infer.sh @@ -0,0 +1,66 @@ +#!/bin/bash + +# Copyright 2017 Xingyu Na +# Apache 2.0 + +. ./path.sh || exit 1; + +if [ $# != 2 ]; then + echo "Usage: $0 " + echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript" + exit 1; +fi + +aishell_audio_dir=$1 +aishell_text=$2/aishell_transcript_v0.8.txt.0604.relabeled.0923.v2 + +#train_dir=data/local/train +#dev_dir=data/local/dev +#test_dir=data/local/test +infer_dir=data/local/infer +tmp_dir=data/local/tmp + +#mkdir -p $train_dir +#mkdir -p $dev_dir +#mkdir -p $test_dir +mkdir -p $infer_dir +mkdir -p $tmp_dir + +# data directory check +if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then + echo "Error: $0 requires two directory arguments" + exit 1; +fi + +# find wav audio file for train, dev and test resp. +find $aishell_audio_dir -iname "*.wav" > $tmp_dir/wav.flist +n=`cat $tmp_dir/wav.flist | wc -l` +# [ $n -ne 141925 ] && \ +# echo Warning: expected 141925 data data files, found $n + +grep -i "wav/infer" $tmp_dir/wav.flist > $infer_dir/wav.flist || exit 1; + +rm -r $tmp_dir + +# Transcriptions preparation +for dir in $infer_dir; do + echo Preparing $dir transcriptions + sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{print $NF}' > $dir/utt.list + sed -e 's/\.wav//' $dir/wav.flist | awk -F '/' '{i=NF-1;printf("%s %s\n",$NF,$i)}' > $dir/utt2spk_all + paste -d' ' $dir/utt.list $dir/wav.flist > $dir/wav.scp_all + utils/filter_scp.pl -f 1 $dir/utt.list $aishell_text > $dir/transcripts.txt + awk '{print $1}' $dir/transcripts.txt > $dir/utt.list + utils/filter_scp.pl -f 1 $dir/utt.list $dir/utt2spk_all | sort -u > $dir/utt2spk + utils/filter_scp.pl -f 1 $dir/utt.list $dir/wav.scp_all | sort -u > $dir/wav.scp + sort -u $dir/transcripts.txt > $dir/text + utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt +done + +mkdir -p data/infer + +for f in spk2utt utt2spk wav.scp text; do + cp $infer_dir/$f data/infer/$f || exit 1; +done + +echo "$0: AISHELL data preparation succeeded" +exit 0; diff --git a/egs/aishell/asr1/run.sh b/egs/aishell/asr1/run.sh index d406715d5d2..45d750b3907 100755 --- a/egs/aishell/asr1/run.sh +++ b/egs/aishell/asr1/run.sh @@ -8,9 +8,10 @@ # general configuration backend=pytorch -stage=0 # start from 0 if you need to start from data preparation -stop_stage=100 -ngpu=1 # number of gpus ("0" uses cpu, otherwise use gpu) +stage=5 # start from 0 if you need to start from data preparation +stop_stage=5 +export CUDA_VISIBLE_DEVICES=0,2,3 +ngpu=3 # number of gpus ("0" uses cpu, otherwise use gpu) debugmode=1 dumpdir=dump # directory to dump full features N=0 # number of minibatches to be used (mainly for debugging). "0" uses all minibatches. @@ -33,8 +34,10 @@ recog_model=model.acc.best # set a model to be used for decoding: 'model.acc.bes n_average=10 # data -data=/export/a05/xna/data -data_url=www.openslr.org/resources/33 +#data=/export/a05/xna/data +#data_url=www.openslr.org/resources/33 + +data=/data/chenghengzhe/kf/dataset/fakes/fake_merged_v2 # exp tag tag="" # tag for managing experiments. @@ -51,11 +54,11 @@ train_set=train_sp train_dev=dev recog_set="dev test" -if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then - echo "stage -1: Data Download" - local/download_and_untar.sh ${data} ${data_url} data_aishell - local/download_and_untar.sh ${data} ${data_url} resource_aishell -fi +# if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then +# echo "stage -1: Data Download" +# local/download_and_untar.sh ${data} ${data_url} data_aishell +# local/download_and_untar.sh ${data} ${data_url} resource_aishell +# fi if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ### Task dependent. You have to make data the following preparation part by yourself. @@ -210,7 +213,7 @@ fi if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then echo "stage 5: Decoding" - nj=32 + nj=40 if [[ $(get_yaml.py ${train_config} model-module) = *transformer* ]]; then recog_model=model.last${n_average}.avg.best average_checkpoints.py --backend ${backend} \ diff --git a/egs/ljspeech/tts1/conf/tuning/train_fastspeech.v1.yaml b/egs/ljspeech/tts1/conf/tuning/train_fastspeech.v1.yaml index f3be22236ab..196449301f6 100644 --- a/egs/ljspeech/tts1/conf/tuning/train_fastspeech.v1.yaml +++ b/egs/ljspeech/tts1/conf/tuning/train_fastspeech.v1.yaml @@ -52,7 +52,7 @@ grad-clip: 1.0 weight-decay: 0.0 patience: 0 epochs: 1000 # 1,000 epochs * 330 batches / 2 accum-grad = 165,000 iters -teacher-model: exp/train_no_dev_pytorch_train_transformer.v1/results/model.last1.avg.best +teacher-model: exp/train_no_dev_pytorch_train_pytorch_transformer.v1/results/model.last1.avg.best # you can download pretrained teacher model from google drive # https://drive.google.com/open?id=1arZAxZOLep-1W5ByQMD1lCX2tEASnw7p diff --git a/egs/ljspeech/tts1/conf/tuning/train_pytorch_transformer.v1.yaml b/egs/ljspeech/tts1/conf/tuning/train_pytorch_transformer.v1.yaml index 7c5512ad253..c7be653b50f 100644 --- a/egs/ljspeech/tts1/conf/tuning/train_pytorch_transformer.v1.yaml +++ b/egs/ljspeech/tts1/conf/tuning/train_pytorch_transformer.v1.yaml @@ -62,7 +62,7 @@ accum-grad: 2 grad-clip: 1.0 weight-decay: 0.0 patience: 0 -epochs: 1000 # 1,000 epochs * 330 batches / 2 accum-grad = 165,000 iters +epochs: 500 # 1,000 epochs * 330 batches / 2 accum-grad = 165,000 iters # other save-interval-epoch: 10 diff --git a/egs/ljspeech/tts1/local/text/cleaners.py b/egs/ljspeech/tts1/local/text/cleaners.py index b99f6eefde1..7fb49245b5f 100644 --- a/egs/ljspeech/tts1/local/text/cleaners.py +++ b/egs/ljspeech/tts1/local/text/cleaners.py @@ -106,8 +106,8 @@ def english_cleaners(text): '''Pipeline for English text, including number and abbreviation expansion.''' text = convert_to_ascii(text) text = lowercase(text) - text = expand_numbers(text) - text = expand_abbreviations(text) + #text = expand_numbers(text) + #text = expand_abbreviations(text) text = expand_symbols(text) text = remove_unnecessary_symbols(text) text = uppercase(text) diff --git a/egs/ljspeech/tts1/run.sh b/egs/ljspeech/tts1/run.sh index eaa2084fe57..5dfd68d6e85 100755 --- a/egs/ljspeech/tts1/run.sh +++ b/egs/ljspeech/tts1/run.sh @@ -8,9 +8,9 @@ # general configuration backend=pytorch -stage=-1 +stage=4 stop_stage=100 -ngpu=1 # number of gpus ("0" uses cpu, otherwise use gpu) +ngpu=4 # number of gpus ("0" uses cpu, otherwise use gpu) nj=32 # numebr of parallel jobs dumpdir=dump # directory to dump full features verbose=0 # verbose option (if set > 0, get more log) @@ -28,9 +28,12 @@ n_shift=256 # number of shift points win_length="" # window length # config files -train_config=conf/train_pytorch_tacotron2.yaml # you can select from conf or conf/tuning. - # now we support tacotron2, transformer, and fastspeech - # see more info in the header of each config. +# you can select from conf or conf/tuning. +# now we support tacotron2, transformer, and fastspeech +# see more info in the header of each config. +train_config=conf/tuning/train_pytorch_transformer.v1.yaml +# train_config=conf/tuning/train_fastspeech.v1.yaml + decode_config=conf/decode.yaml # decoding related @@ -52,20 +55,21 @@ set -e set -u set -o pipefail +corpus_path="/data/nipeng/TTS/data/fake_lj_nosil" train_set="train_no_dev" dev_set="dev" eval_set="eval" -if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then - echo "stage -1: Data Download" - local/download.sh ${db_root} -fi +# if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then +# echo "stage -1: Data Download" +# local/download.sh ${db_root} +# fi if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ### Task dependent. You have to make data the following preparation part by yourself. ### But you can utilize Kaldi recipes in most cases echo "stage 0: Data preparation" - local/data_prep.sh ${db_root}/LJSpeech-1.1 data/train + local/data_prep.sh ${corpus_path} data/train utils/validate_data_dir.sh --no-feats data/train fi diff --git a/tools/Makefile b/tools/Makefile index 8c88f218678..78129894166 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -1,15 +1,15 @@ # If a Python interpreter is specified, then creates a virtualenv from it # PYTHON := /usr/bin/python3.7 -PYTHON := +PYTHON := /root/anaconda3/envs/np-py36-tts/bin/python3 # The python version installed in the conda setup # NOTE(kan-bayashi): Use 3.7.3 to avoid sentencepiece installation error -PYTHON_VERSION := 3.7.3 +PYTHON_VERSION := 3.6.8 CUPY_VERSION := 6.0.0 -CUDA_VERSION := 10.0 +CUDA_VERSION := 9.0 # PyTorch version: 0.4.1 or 1.0.0 or 1.0.1 -TH_VERSION := 1.0.1 +TH_VERSION := 1.0.0 # Use a prebuild Kaldi to omit the installation -KALDI := +KALDI := /data/nipeng/kaldi WGET := wget --tries=3 # Both Miniconda2/3 can install any Python versions From 0d784e384aa06b12e41180e8191493dffdbe4169 Mon Sep 17 00:00:00 2001 From: nipeng Date: Thu, 24 Oct 2019 12:33:36 +0800 Subject: [PATCH 02/23] update --- egs/aishell/asr1/infer.sh | 3 ++- egs/aishell/asr1/local/aishell_data_prep.sh | 2 +- egs/aishell/asr1/run.sh | 8 ++++---- egs/aishell/asr1/test.py | 22 +++++++++++++++++++++ 4 files changed, 29 insertions(+), 6 deletions(-) create mode 100644 egs/aishell/asr1/test.py diff --git a/egs/aishell/asr1/infer.sh b/egs/aishell/asr1/infer.sh index b9cbb023099..731d21832d0 100755 --- a/egs/aishell/asr1/infer.sh +++ b/egs/aishell/asr1/infer.sh @@ -36,7 +36,8 @@ n_average=10 #data=/export/a05/xna/data #data_url=www.openslr.org/resources/33 -data=/data/nipeng/2019-10-07-aishell +# data=/data/nipeng/2019-10-07-aishell +data=/data/nipeng/2019-1001-1020-aishell # exp tag tag="" # tag for managing experiments. diff --git a/egs/aishell/asr1/local/aishell_data_prep.sh b/egs/aishell/asr1/local/aishell_data_prep.sh index 700afa3dee9..a4e411f175b 100755 --- a/egs/aishell/asr1/local/aishell_data_prep.sh +++ b/egs/aishell/asr1/local/aishell_data_prep.sh @@ -12,7 +12,7 @@ if [ $# != 2 ]; then fi aishell_audio_dir=$1 -aishell_text=$2/aishell_transcript_v0.8.txt.0604.relabeled.0923.v2 +aishell_text=$2/aishell_transcript_v0.8.txt.add10w.20190918.tune_high_medium_lv.v2 train_dir=data/local/train dev_dir=data/local/dev diff --git a/egs/aishell/asr1/run.sh b/egs/aishell/asr1/run.sh index 45d750b3907..77119cf5694 100755 --- a/egs/aishell/asr1/run.sh +++ b/egs/aishell/asr1/run.sh @@ -8,10 +8,10 @@ # general configuration backend=pytorch -stage=5 # start from 0 if you need to start from data preparation -stop_stage=5 -export CUDA_VISIBLE_DEVICES=0,2,3 -ngpu=3 # number of gpus ("0" uses cpu, otherwise use gpu) +stage=0 # start from 0 if you need to start from data preparation +stop_stage=100 +# export CUDA_VISIBLE_DEVICES=0,2,3 +ngpu=4 # number of gpus ("0" uses cpu, otherwise use gpu) debugmode=1 dumpdir=dump # directory to dump full features N=0 # number of minibatches to be used (mainly for debugging). "0" uses all minibatches. diff --git a/egs/aishell/asr1/test.py b/egs/aishell/asr1/test.py new file mode 100644 index 00000000000..bfb1cc33981 --- /dev/null +++ b/egs/aishell/asr1/test.py @@ -0,0 +1,22 @@ +# -*- coding: utf-8 -*- + + +import codecs + + + +def get_infer_result(): + path = "/data/nipeng/TTS/espnet/egs/aishell/asr1/exp/train_sp_pytorch_train/decode_infer_decode_lm/hyp.trn" + fout = codecs.open("infer.txt", "w") + with codecs.open(path) as f: + for line in f: + t, a = line.split("(") + text = t.replace(" ", "") + aid = a.split("-")[0] + aname = aid + ".wav" + fout.write("%s\t%s\n" %(aname, text)) + + +if __name__ == "__main__": + print("ok") + get_infer_result() From 878eaef9934bcf6e12e29c3132542a6e4a681503 Mon Sep 17 00:00:00 2001 From: root Date: Thu, 24 Oct 2019 16:30:44 +0800 Subject: [PATCH 03/23] =?UTF-8?q?=E4=BF=AE=E6=94=B9python,kaldi=E9=BB=98?= =?UTF-8?q?=E8=AE=A4=E8=B7=AF=E5=BE=84?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- tools/Makefile | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tools/Makefile b/tools/Makefile index 78129894166..11173755732 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -1,6 +1,7 @@ # If a Python interpreter is specified, then creates a virtualenv from it # PYTHON := /usr/bin/python3.7 -PYTHON := /root/anaconda3/envs/np-py36-tts/bin/python3 +# PYTHON := /root/anaconda3/envs/np-py36-tts/bin/python3 +PYTHON := # The python version installed in the conda setup # NOTE(kan-bayashi): Use 3.7.3 to avoid sentencepiece installation error PYTHON_VERSION := 3.6.8 @@ -9,7 +10,8 @@ CUDA_VERSION := 9.0 # PyTorch version: 0.4.1 or 1.0.0 or 1.0.1 TH_VERSION := 1.0.0 # Use a prebuild Kaldi to omit the installation -KALDI := /data/nipeng/kaldi +# KALDI := /data/nipeng/kaldi +KALDI := WGET := wget --tries=3 # Both Miniconda2/3 can install any Python versions From f44d6fdc2590803ab5bef15cc95f9703d07fd28c Mon Sep 17 00:00:00 2001 From: nipeng Date: Thu, 24 Oct 2019 22:25:43 +0800 Subject: [PATCH 04/23] add: conda config channels --- tools/Makefile | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tools/Makefile b/tools/Makefile index 11173755732..22951a259c2 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -63,6 +63,9 @@ miniconda.sh: test -f miniconda.sh || $(WGET) $(CONDA_URL) -O miniconda.sh venv: miniconda.sh test -d $(PWD)/venv || bash miniconda.sh -b -p $(PWD)/venv + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ . venv/bin/activate && conda update -y conda . venv/bin/activate && conda install -y python=$(PYTHON_VERSION) . venv/bin/activate && conda info -a From 964803fc52636922e8c1be4dd145a9f44cec13f4 Mon Sep 17 00:00:00 2001 From: nipeng Date: Fri, 25 Oct 2019 09:16:09 +0800 Subject: [PATCH 05/23] add: makefile --- tools/Makefile | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tools/Makefile b/tools/Makefile index 22951a259c2..44f5552cf3c 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -63,10 +63,11 @@ miniconda.sh: test -f miniconda.sh || $(WGET) $(CONDA_URL) -O miniconda.sh venv: miniconda.sh test -d $(PWD)/venv || bash miniconda.sh -b -p $(PWD)/venv + . venv/bin/activate && conda update -y conda . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ - . venv/bin/activate && conda update -y conda + . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y python=$(PYTHON_VERSION) . venv/bin/activate && conda info -a espnet.done: venv From d9262d8d109516f1947ec32bf0bdf8a11dfba066 Mon Sep 17 00:00:00 2001 From: nipeng Date: Fri, 25 Oct 2019 09:38:55 +0800 Subject: [PATCH 06/23] add: pip index --- tools/Makefile | 1 + 1 file changed, 1 insertion(+) diff --git a/tools/Makefile b/tools/Makefile index 44f5552cf3c..61c4b850a63 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -56,6 +56,7 @@ venv: espnet.done: venv . venv/bin/activate; pip install pip --upgrade . venv/bin/activate; pip install -e .. + . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ . venv/bin/activate; pip install torch==$(TH_VERSION) touch espnet.done else From 9634530d2e853e775e045add82d28029fdf23fa2 Mon Sep 17 00:00:00 2001 From: nipeng Date: Fri, 25 Oct 2019 11:23:22 +0800 Subject: [PATCH 07/23] add: pip --- tools/Makefile | 1 + 1 file changed, 1 insertion(+) diff --git a/tools/Makefile b/tools/Makefile index 61c4b850a63..d75297666bd 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -54,6 +54,7 @@ ifneq ($(strip $(PYTHON)),) venv: test -d venv || virtualenv -p $(PYTHON) venv espnet.done: venv + . venv/bin/activate; pip install -U pip . venv/bin/activate; pip install pip --upgrade . venv/bin/activate; pip install -e .. . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ From b25af285cf702dfe2b2661a127dc470f65ee4f8b Mon Sep 17 00:00:00 2001 From: nipeng Date: Fri, 25 Oct 2019 14:52:27 +0800 Subject: [PATCH 08/23] =?UTF-8?q?=E4=BF=AE=E6=94=B9python=E7=89=88?= =?UTF-8?q?=E6=9C=AC?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- tools/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/Makefile b/tools/Makefile index d75297666bd..6810c9a2642 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -4,7 +4,7 @@ PYTHON := # The python version installed in the conda setup # NOTE(kan-bayashi): Use 3.7.3 to avoid sentencepiece installation error -PYTHON_VERSION := 3.6.8 +PYTHON_VERSION := 3.7.3 CUPY_VERSION := 6.0.0 CUDA_VERSION := 9.0 # PyTorch version: 0.4.1 or 1.0.0 or 1.0.1 From e669ea4015b1f57d104b2e3736314b093b48a07b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E8=92=99?= Date: Fri, 25 Oct 2019 16:59:59 +0800 Subject: [PATCH 09/23] pip --- tools/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/Makefile b/tools/Makefile index 6810c9a2642..c3250242d57 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -58,7 +58,7 @@ espnet.done: venv . venv/bin/activate; pip install pip --upgrade . venv/bin/activate; pip install -e .. . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ - . venv/bin/activate; pip install torch==$(TH_VERSION) + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ torch==$(TH_VERSION) touch espnet.done else miniconda.sh: From f1c55f12c074426d330757b4f16bbf0fbd3745b6 Mon Sep 17 00:00:00 2001 From: nipeng Date: Fri, 25 Oct 2019 17:27:14 +0800 Subject: [PATCH 10/23] add: conda channels --- tools/Makefile | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tools/Makefile b/tools/Makefile index c3250242d57..d3c6a38ede3 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -73,6 +73,10 @@ venv: miniconda.sh . venv/bin/activate && conda install -y python=$(PYTHON_VERSION) . venv/bin/activate && conda info -a espnet.done: venv + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ + . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y $(CONDA_PYTORCH) -c pytorch . venv/bin/activate && pip install -e .. touch espnet.done From 19d3e85d00e54cb5e7df6a9a1bdc0942a719161a Mon Sep 17 00:00:00 2001 From: nipeng Date: Sat, 26 Oct 2019 09:29:29 +0800 Subject: [PATCH 11/23] =?UTF-8?q?=E5=A2=9E=E5=8A=A0infer?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- docker/infer.sh | 163 ++++++++++++++++++ .../asr1/local/aishell_data_prep_infer.sh | 1 + 2 files changed, 164 insertions(+) create mode 100755 docker/infer.sh diff --git a/docker/infer.sh b/docker/infer.sh new file mode 100755 index 00000000000..e7b6ec914a1 --- /dev/null +++ b/docker/infer.sh @@ -0,0 +1,163 @@ +#!/bin/bash + +docker_gpu=0 +docker_egs= +docker_folders= +docker_cuda=10.0 +docker_user=true +docker_env= +docker_cmd= +docker_os=u18 + +while test $# -gt 0 +do + case "$1" in + -h) echo "Usage: `basename $0` [-h] docker_gpu docker_egs docker_folders options" + exit 0;; + --help) echo "Usage: `basename $0` [-h] ] docker_gpu docker_egs docker_folders options" + exit 0;; + --docker*) ext=${1#--} + frombreak=true + for i in _ {a..z} {A..Z}; do + for var in `eval echo "\\${!${i}@}"`; do + if [ "$var" == "$ext" ]; then + eval ${ext}=$2 + frombreak=false + break 2 + fi + done + done + if ${frombreak} ; then + echo "bad option $1" + exit 1 + fi + ;; + --*) break + ;; + esac + shift + shift +done + +if [ -z "${docker_egs}" ]; then + echo "Select an example to work with from the egs folder." + exit 1 +fi + +from_tag="cpu" +if [ ! "${docker_gpu}" == "-1" ]; then + if [ -z "${docker_cuda}" ]; then + # If the docker_cuda is not set, the program will automatically + # search the installed version with default configurations (apt) + docker_cuda=$( nvcc -V | grep release ) + docker_cuda=${docker_cuda#*"release "} + docker_cuda=${docker_cuda%,*} + fi + # After search for your cuda version, if the variable docker_cuda is empty the program will raise an error + if [ -z "${docker_cuda}" ]; then + echo "CUDA was not found in your system. Use CPU image or install NVIDIA-DOCKER, CUDA and NVCC for GPU image." + exit 1 + else + from_tag="gpu-cuda${docker_cuda}-cudnn7" + fi +fi + +if [ ! -z "${docker_os}" ]; then + from_tag="${from_tag}-${docker_os}" +fi + +# Check if image exists in the system and download if required +docker_image=$( docker images -q espnet/espnet:${from_tag} ) +if ! [[ -n ${docker_image} ]]; then + docker pull espnet/espnet:${from_tag} +fi + +if [ ${docker_user} = true ]; then + # Build a container with the user account + container_tag="${from_tag}-user-${HOME##*/}" + docker_image=$( docker images -q espnet/espnet:${container_tag} ) + if ! [[ -n ${docker_image} ]]; then + echo "Building docker image..." + build_args="--build-arg FROM_TAG=${from_tag}" + build_args="${build_args} --build-arg THIS_USER=${HOME##*/}" + build_args="${build_args} --build-arg THIS_UID=${UID}" + + echo "Now running docker build ${build_args} -f prebuilt/Dockerfile -t espnet/espnet:${container_tag} ." + (docker build ${build_args} -f prebuilt/Dockerfile -t espnet/espnet:${container_tag} .) || exit 1 + fi +else + container_tag=${from_tag} +fi + +echo "Using image espnet/espnet:${container_tag}." + +this_time="$(date '+%Y%m%dT%H%M')" +if [ "${docker_gpu}" == "-1" ]; then + cmd0="docker" + container_name="espnet_cpu_${this_time}" +else + # --rm erase the container when the training is finished. + cmd0="NV_GPU='${docker_gpu}' nvidia-docker" + container_name="espnet_gpu${docker_gpu//,/_}_${this_time}" +fi + +cd .. + +vols="-v /data/mfs:/data -v ${PWD}/egs:/espnet/egs -v ${PWD}/espnet:/espnet/espnet -v ${PWD}/test:/espnet/test -v ${PWD}/utils:/espnet/utils" +if [ ! -z "${docker_folders}" ]; then + docker_folders=$(echo ${docker_folders} | tr "," "\n") + for i in ${docker_folders[@]} + do + vols=${vols}" -v $i:$i"; + done +fi + +cmd1="cd /espnet/egs/${docker_egs}" +if [ ! -z "${docker_cmd}" ]; then + cmd2="./${docker_cmd} $@" +else + cmd2="./run.sh $@" +fi + +if [ ${docker_user} = false ]; then + # Required to access to the folder once the training if finished in root access + cmd2="${cmd2}; chmod -R 777 /espnet/egs/${docker_egs}" +fi + +cmd="${cmd1}; ${cmd2}" +this_env="" +if [ ! -z "${docker_env}" ]; then + docker_env=$(echo ${docker_env} | tr "," "\n") + for i in ${docker_env[@]} + do + this_env="-e $i ${this_env}" + done +fi + +if [ ! -z "${HTTP_PROXY}" ]; then + this_env="${this_env} -e 'HTTP_PROXY=${HTTP_PROXY}'" +fi + +if [ ! -z "${http_proxy}" ]; then + this_env="${this_env} -e 'http_proxy=${http_proxy}'" +fi + +cmd="${cmd0} run -i -v /data/mfs:/data/mfs --rm ${this_env} --name ${container_name} ${vols} espnet/espnet:${container_tag} /bin/bash -c '${cmd}'" + + +trap ctrl_c INT + +function ctrl_c() { + echo "** Kill docker container ${container_name}" + docker rm -f ${container_name} +} + +echo "Executing application in Docker" +echo ${cmd} +eval ${cmd} & +PROC_ID=$! + +while kill -0 "$PROC_ID" 2> /dev/null; do + sleep 1 +done +echo "`basename $0` done." diff --git a/egs/aishell/asr1/local/aishell_data_prep_infer.sh b/egs/aishell/asr1/local/aishell_data_prep_infer.sh index f4b75d102a7..c598ed5fc13 100755 --- a/egs/aishell/asr1/local/aishell_data_prep_infer.sh +++ b/egs/aishell/asr1/local/aishell_data_prep_infer.sh @@ -29,6 +29,7 @@ mkdir -p $tmp_dir # data directory check if [ ! -d $aishell_audio_dir ] || [ ! -f $aishell_text ]; then echo "Error: $0 requires two directory arguments" + echo ${aishell_audio_dir}, ${aishell_text} exit 1; fi From 0c3940894cc049119b3362135db7234d7a3c9fd0 Mon Sep 17 00:00:00 2001 From: nipeng Date: Mon, 28 Oct 2019 18:26:36 +0800 Subject: [PATCH 12/23] change trans file name --- egs/aishell/asr1/asr_pipline.py | 453 ++++++++++++++++++ egs/aishell/asr1/asr_pipline.sh | 34 ++ .../asr1/local/aishell_data_prep_infer.sh | 2 +- 3 files changed, 488 insertions(+), 1 deletion(-) create mode 100644 egs/aishell/asr1/asr_pipline.py create mode 100644 egs/aishell/asr1/asr_pipline.sh diff --git a/egs/aishell/asr1/asr_pipline.py b/egs/aishell/asr1/asr_pipline.py new file mode 100644 index 00000000000..50bb04b88e4 --- /dev/null +++ b/egs/aishell/asr1/asr_pipline.py @@ -0,0 +1,453 @@ +# -*- coding: utf-8 -*- + +import sys + +if sys.version[0] == '2': + reload(sys) + sys.setdefaultencoding("utf-8") + +import os +import time +import json +import redis +import jieba +import codecs +import shutil +import hashlib +import tempfile +import argparse +import subprocess +from collections import defaultdict +from functools import partial +from concurrent.futures import ProcessPoolExecutor +from multiprocessing import cpu_count +from pydub import AudioSegment +from kafka import KafkaConsumer, KafkaProducer +from kafka.errors import CommitFailedError +from elasticsearch import Elasticsearch, helpers + + +def norm_aishell_data(indir): + # /tmp/asr_9hmlx3zl + outdir = indir + "_norm" + for path in os.listdir(indir): + fname = path.split(".")[0] + os.system("mkdir -p %s/data_aishell/wav/infer/%s" %(outdir, fname)) + cmd = "cp %s/%s %s/data_aishell/wav/infer/%s/." %(indir, path, outdir, fname) + os.system("cp %s/%s %s/data_aishell/wav/infer/%s/." %(indir, path, outdir, fname)) + + os.system("mkdir -p %s/data_aishell/transcript" %outdir) + outfile = "%s/data_aishell/transcript/aishell_transcript.txt" %outdir + fout = codecs.open(outfile, "w") + for path in os.listdir(indir): + fname = path.split(".")[0] + fout.write("%s %s\n" %(fname, "哈哈")) + return outdir + + +def gen_sp_wav_and_get_path_one(wav_temppath, audio_id, sound, item, k): + cut_wav_name = "%s_%s_%s_%s.wav" % (audio_id, item['start'], item['end'], k) + save_cut_path = os.path.join(wav_temppath, cut_wav_name) + sp_wav = sound[int(item['start']):int(item['end'])] + if sp_wav.frame_rate != 8000: + sp_wav = sp_wav.set_frame_rate(8000) + sp_wav.export(save_cut_path, format="wav") + print("==3", save_cut_path) + return save_cut_path + + +def get_parser(): + parser = argparse.ArgumentParser(description='语音识别主函数参数') + parser.add_argument("-kh", "--kafka-host", + default="192.168.40.22:9090,192.168.40.19:9090,192.168.40.59:9090,192.168.40.60:9090,192.168.40.61:9090", + required=True, + help="kafka host:port 集群") + + parser.add_argument("-sct", "--seg-consumer-topics", + default="sp_vad_topic1", + help="输入kafka的topic") + + parser.add_argument("-ikg", "--seg-consumer-groupid", + default="asr_group1", + help="消费kafka的groupid") + + parser.add_argument("-stm", "--session-timeout-ms", + default=60000, + type=int, + help="消费kafka的心跳超时时间") + + parser.add_argument("-saor", "--seg-auto-offset-reset", + default="largest", + help="重置偏移量,earliest移到最早的可用消息,latest最新的消息, 默认为largest,即是latest.消费者消费类型参数:{'smallest': 'earliest', 'largest': 'latest'}") + + parser.add_argument("-apt", "--asr-producer-topics", + default="asr_topic1", + help="输入kafka的topic") + + parser.add_argument("-fp", "--father-path", + default="/data/mfs/k8s/speech_pipeline/sad", + help="切分来源音频存放父目录") + + parser.add_argument("-wp", "--hkust-path", + default="/home/app/asr_pipline/kaldi/egs/hkust/s5_iqianjin", + help="hkust的绝对目录,即kaldi的hkust目录") + + parser.add_argument("-nj", "--num-job", + type=int, default=10, + help="hkust num job default 10") + + parser.add_argument("-nw", "--num-workers", + type=int, default=18, + help="multiprocess number of workers") + + parser.add_argument("-cg", "--consumer-gap", + type=int, default=2, + help="kafka consumer msg num") + + parser.add_argument("-ptm", "--poll-timeout-ms", + type=int, default=60000, + help="") + + parser.add_argument('--is-comp', help='1为补数,0为不进入补数筛选逻辑', default=0, type=int) + args = parser.parse_args() + return args + + +class ASR(object): + def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, + session_timeout_ms=60000, seg_auto_offset_reset="largest", + asr_producer_topics="asr_topic1", num_job=10, is_comp=0, + poll_timeout_ms=60000, consumer_gap=None, num_workers=cpu_count()): + """ + :param kafka_servers: kafka host:port + :param seg_consumer_topics: 切分的消费者topic + :param seg_consumer_groupid: 消费者group id + :param session_timeout_ms: 心跳超时时间 + :param seg_auto_offset_reset: 重置偏移量, earliest移到最早的可用消息, latest最新的消息, 默认为largest,即是latest. + 源码定义: {'smallest': 'earliest', 'largest': 'latest'} + :param asr_producer_topics: 语音是被的生产者topic,默认值:asr_topic1 + :param num_job: 语音识别的线程数 + :param is_comp: 是否为补数逻辑,是的话会进入时间判断和es数据判断 + """ + self.kafka_servers = kafka_servers + self.seg_consumer_groupid = seg_consumer_groupid + self.session_timeout_ms = session_timeout_ms + self.seg_auto_offset_reset = seg_auto_offset_reset + self.seg_consumer_topics = seg_consumer_topics + self.num_job = num_job + self.poll_timeout_ms = poll_timeout_ms + self.consumer_gap = consumer_gap + self.num_workers = num_workers + + self._get_from_client() + + # 语音识别结果的kafka生产者 + self.to_client = KafkaProducer(bootstrap_servers=kafka_servers, # kafka host:port + compression_type="gzip", + max_request_size=1024 * 1024 * 20) + self.asr_producer_topics = asr_producer_topics # ASR生产者topic + + self.is_comp = is_comp + if is_comp: + self.comp_num = 0 + self.prop = {"HOST1": "192.168.40.37", + "HOST2": "192.168.40.38", + "HOST3": "192.168.40.39", + "PORT": "9200", + "DOC_TYPE": "kf_infobird_call"} + + self.es = Elasticsearch(hosts=[{'host': self.prop['HOST1'], 'port': self.prop['PORT']}, + {'host': self.prop['HOST2'], 'port': self.prop['PORT']}, + {'host': self.prop['HOST3'], 'port': self.prop['PORT']}]) + + def _get_from_client(self): + # 消费者切分好的音频kafka消费者 + self.from_client = KafkaConsumer(bootstrap_servers=self.kafka_servers, # kafka host:port + group_id=self.seg_consumer_groupid, # 消费者group id + session_timeout_ms=self.session_timeout_ms, # 设置心跳超时时间 + enable_auto_commit=False, # 是否自动提交 + auto_offset_reset=self.seg_auto_offset_reset) # 消费重置偏移量 + self.from_client.subscribe(self.seg_consumer_topics) # 切分的消费者topic + + def asr_pipline_from_kafka(self, father_path): + """ + 获取kafka的数据流，并进行识别，合并，标点，存入es + :param father_path: 切分来源音频存放父目录 + :return: + """ + # redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") + while True: + if not self.from_client: + self._get_from_client() + + tp_msgs = self.from_client.poll(timeout_ms=self.poll_timeout_ms, + max_records=self.consumer_gap) + msgs = [] + for tp, _msgs in tp_msgs.items(): + msgs.extend(_msgs) + print(len(msgs)) + self.batch_asr_pipline(father_path, msgs) + break + + def batch_asr_pipline(self, father_path, msgs): + """ + 单个kafka消息消费: + 1)从kafka消息中提取value信息; + 2)根据msg.value信息从切分好的数据源复制到目标位置; + 3)进行语音识别,放进merge_dict里; + 4)写入kafka新的topic; + 7)删除临时音频的文件夹和语音识别结果的临时文件. + :param father_path: 切分来源音频存放父目录; + :param msg: 从kafka获取的完整消息; + :return: + """ + flag = False # flag 是否语音识别成功并存入kafka + wav_temppath = tempfile.mkdtemp(prefix="asr_") + batch_wav_lst = [] + batch_voice_data = {} + batch_merge_dict = None + +# try: + redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") + for msg in msgs: + if msg is not None: + audio_id = json.loads(msg.value).get('audio_id', '') + print("==2", audio_id) + if not redis_client.get('asr_espnet_' + str(audio_id)): + # step1: 从kafka消息中提取value信息 + voice_data = json.loads(msg.value) + voice_data['start_asr'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + batch_voice_data[audio_id] = voice_data # 后面继续往 voice_data 追加数据 + # step2：根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav + batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data) + # step3: 语音识别, 并获取merge的text + # batch results + print("==4", wav_temppath) + wav_normpath = norm_aishell_data(wav_temppath) + merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} + #batch_merge_dict = self.split_merge_dict(merge_dict) +# except Exception as e: +# print("bebefore commit to kafka error log: %s, msg: %s" % (e, "")) +# finally: +# # step7: 删除临时音频的文件夹和语音识别结果的临时文件 +# #shutil.rmtree(wav_temppath) +# pass + return + + for audio_id, voice_data in batch_voice_data.items(): + # step4: 写入kafka新的topic + try: + voice_data["merge_dict"] = batch_merge_dict[audio_id] + voice_data['end_asr'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + flag = self._kafka_producers(voice_data) if voice_data.get("merge_dict", {}) else False + voice_data['step4_end_time'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + print("asr_output:", json.dumps(voice_data, ensure_ascii=True)) + except Exception as e: + print("before commit to kafka error log: %s, msg: %s" % (e, "")) + try: + self.from_client.commit() + except CommitFailedError: + print('commit error!') + for audio_id, _ in batch_voice_data.items(): + redis_client.set('asr_espnet_' + str(audio_id), 1, ex=24 * 3600) + + return flag + + def split_merge_dict(self, merge_dict): + """ + 拆分merge_dict + :param merge_dict: {"A_00*_**.wav": "123", + "A_**_***.wav": "321", + "B_**_***.wav": "423"} + """ + split_dict = defaultdict(dict) + for wav_file, text in merge_dict.items(): + audio_id = wav_file.split("_")[0] + split_dict[audio_id][wav_file] = text + return split_dict + + def filter_conditions(self, voice_data): + """ + 筛选出不需要补数的数据,并抛出异常; + :param voice_data: + :return: + """ + end_sad = voice_data.get("end_sad", "0000-00-00 00:00:00") + audio_id = voice_data.get("audio_id", None) + body = { + "query": {"bool": {"must": [{"term": {"audio_id.keyword": audio_id}}], "must_not": [], "should": []}}, + "from": 0, "size": 10, "sort": [], "aggs": {}} + if end_sad < "2019-07-11 12:00:00" or end_sad > "2019-07-11 21:00:00": + raise RuntimeError('This one\'s end_sad not in time window !') + query = self.es.search(index="asr_text", + doc_type=self.prop['DOC_TYPE'], + scroll='5m', + timeout='30s', + body=body) + results = query['hits']['hits'] # es查询出的结果第一页 + total = query['hits']['total'] # es查询出的结果总量 + print("audio_id:", audio_id, "search num:", total, ",end_sad", end_sad) + if total > 0: + line = results[0]["_source"] + if line.get("cut_text"): + raise RuntimeError('This id already exists and already asr !') + else: + segments = voice_data.get("segments", {}) + for k, v in segments.items(): + if k in ["gk", "kf"] and len(v) > 0: + break + else: + raise RuntimeError('This data segments is null !') + self.comp_num += 1 + + def _copy_wav(self, father_path, voice_data, temp_path): + """ + step2: 根据msg信息从切分好的数据源复制到目标位置 + :param father_path: 切分来源音频存放父目录 + :param voice_data: kafka来源信息中的包含切分音频的信息 + :param temp_path: 音频复制目标目录 + :return: wav_lst: + """ + source = voice_data.get("source", "") + date = voice_data.get("date", "") + audio_id = voice_data.get("audio_id", "") + + value_dict = voice_data.get("segments", {}) + wav_lst = [] # 音频list + + for chn, split_time in value_dict.items(): + for si in split_time: + split_path = "%s_%s_%s_%s.wav" % (audio_id, si.get("start"), si.get("end"), chn) + wav_path = os.path.join(father_path, source, date, audio_id, split_path) + shutil.copy(wav_path, temp_path) + wav_lst.append(wav_path) + return wav_lst + + def gen_sp_wav_and_get_path(self, father_path, wav_temppath, voice_data): + """ + step2: 根据msg信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav + :param father_path: 音频来源目录,结果存储 + :param wav_temppath: kafka来源信息中的包含切分音频的信息 + :return: wav_lst: + """ + wav_lst = [] # wav 音频存储 + source = voice_data["source"] + code = None if source != 'infobird' else 'pcm_s16le' + date = voice_data["date"] + audio_id = voice_data["audio_id"] + wav_father_path = os.path.join(father_path, source, date) # /data/mfs/k8s/speech_pipeline/raw/{source}/{date} + for k, v in voice_data['segments'].items(): + if k in ["gk", "kf"] and len(v) > 0: + wav_name = "%s_%s.wav" % (audio_id, k) + raw_wav_path = os.path.join(wav_father_path, wav_name) + sound = AudioSegment.from_file(raw_wav_path, codec=code, format="wav") + for item in v: + cut_wav_name = "%s_%s_%s_%s.wav" % (audio_id, item['start'], item['end'], k) + save_cut_path = os.path.join(wav_temppath, cut_wav_name) + sp_wav = sound[int(item['start']):int(item['end'])] + if sp_wav.frame_rate != 8000: + sp_wav = sp_wav.set_frame_rate(8000) + sp_wav.export(save_cut_path, format="wav") + wav_lst.append(save_cut_path) + return wav_lst + + def gen_sp_wav_and_get_path_mp(self, father_path, wav_temppath, batch_voice_data): + """ + multiprocess generator + step2-1: 根据msg信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav + :param father_path: 音频来源目录,结果存储 + :param wav_temppath: kafka来源信息中的包含切分音频的信息 + :return: wav_lst: + """ + p = ProcessPoolExecutor(max_workers=self.num_workers) #不填则默认为cpu的个数 + result = [] + wav_lst = [] # wav 音频存储 + for _, voice_data in batch_voice_data.items(): + source = voice_data["source"] + code = None if source != 'infobird' else 'pcm_s16le' + date = voice_data["date"] + audio_id = voice_data["audio_id"] + wav_father_path = os.path.join(father_path, source, date) # /data/mfs/k8s/speech_pipeline/raw/{source}/{date} + for k, v in voice_data['segments'].items(): + if k in ["gk", "kf"] and len(v) > 0: + wav_name = "%s_%s.wav" % (audio_id, k) + raw_wav_path = os.path.join(wav_father_path, wav_name) + sound = AudioSegment.from_file(raw_wav_path, codec=code, format="wav") + for item in v: + # print(item, type(item)) + obj = p.submit(partial(gen_sp_wav_and_get_path_one, wav_temppath, audio_id, sound, item, k)) + result.append(obj) + p.shutdown() + res = [obj.result() for obj in result] + return res + + def _asr_cmd(self, wav_path): + """ + step3: 语音识别, 并获取merge的text + :param hkust_path: hkust的绝对目录,即kaldi的hkust目录 + :param wav_path: 需识别音频存放目录 + :return: + """ + decode_cmd = "./infer.sh {}".format(wav_path) + print(decode_cmd) + os.system(decode_cmd) + ## return merge_dict + + def _kafka_producers(self, voice_data): + """ + step6: 将语音识别好的结果存入kafka中 + :param voice_data: 来源kafka音频结果与语音识别结果结合体,从中获取相关信息，包含识别结果 + :return: + """ + flag = False + try: + audio_id = voice_data.get("audio_id", "") + asr_model = voice_data.get("asr_model", "") + k = self._create_id_by_input(audio_id + asr_model).encode("utf8") + v = json.dumps(voice_data).encode("utf8") + self.to_client.send(topic=self.asr_producer_topics, key=k, value=v) + flag = True + except Exception as e: + # logger.error("error: %s, voice_data: %s" % (e, voice_data)) + print("error: %s, voice_data: %" % (e, json.dumps(voice_data, ensure_ascii=True))) + finally: + return flag + + def _create_id_by_input(self, id=""): + """ + 依据输入与时间生成一个唯一的md5的id + :param x: + :return: + """ + x = str(id).encode("utf8") + m = hashlib.md5(x) + return m.hexdigest() + + +if __name__ == '__main__': + args = get_parser() + kafka_host, session_timeout_ms = args.kafka_host, args.session_timeout_ms + seg_consumer_topics, seg_auto_offset_reset = args.seg_consumer_topics, args.seg_auto_offset_reset # 消费切分结果的消费者参数 + seg_consumer_groupid, asr_producer_topics = args.seg_consumer_groupid, args.asr_producer_topics # 语音识别生产者参数 + father_path, hkust_path = args.father_path, args.hkust_path # 音频与模型相关参数 + num_job = args.num_job # 线程数 + is_comp = args.is_comp # 是否走补数过滤流程 + poll_timeout_ms = args.poll_timeout_ms + consumer_gap = args.consumer_gap + num_workers = args.num_workers + + # python2 asr_pipline.py --kafka-host 192.168.40.22:9090,192.168.40.19:9090,192.168.40.59:9090,192.168.40.60:9090,192.168.40.61:9090 \ + # --seg_consumer_topics sp_vad_topic1 \ + # --seg_consumer_groupid asr_group1 \ + # --seg-auto-offset-reset largest \ + # --father-path /data/mfs/k8s/speech_pipeline/sad \ + # --hkust-path /home/app/asr_pipline/kaldi/egs/hkust/s5_iqianjin + + if type(seg_consumer_topics) is str: + seg_consumer_topics = [seg_consumer_topics, ] + asr = ASR(kafka_servers=kafka_host, seg_consumer_topics=seg_consumer_topics, session_timeout_ms=session_timeout_ms, + seg_consumer_groupid=seg_consumer_groupid, seg_auto_offset_reset=seg_auto_offset_reset, + asr_producer_topics=asr_producer_topics, num_job=num_job, is_comp=is_comp, + poll_timeout_ms=poll_timeout_ms, consumer_gap=consumer_gap, num_workers=num_workers) + + asr.asr_pipline_from_kafka(father_path) diff --git a/egs/aishell/asr1/asr_pipline.sh b/egs/aishell/asr1/asr_pipline.sh new file mode 100644 index 00000000000..52af73ac321 --- /dev/null +++ b/egs/aishell/asr1/asr_pipline.sh @@ -0,0 +1,34 @@ +#!/usr/bin/bash +#vim /etc/hosts +## Entries added by HostAliases. +#192.168.176.181 daasoffline1.kafka.dc.puhuifinance.com +#192.168.176.182 daasoffline2.kafka.dc.puhuifinance.com +#192.168.176.183 daasoffline3.kafka.dc.puhuifinance.com +#192.168.176.184 daasoffline4.kafka.dc.puhuifinance.com +#192.168.176.185 daasoffline5.kafka.dc.puhuifinance.com + +# test +# 测试需要在代码内修改生产则kafka host +python3 asr_pipline.py \ + --kafka-host daasoffline1.kafka.dc.puhuifinance.com:6667,daasoffline2.kafka.dc.puhuifinance.com:6667,daasoffline3.kafka.dc.puhuifinance.com:6667,daasoffline4.kafka.dc.puhuifinance.com:6667,daasoffline5.kafka.dc.puhuifinance.com:6667 \ + --seg-consumer-topics sp_sad_topic \ + --seg-consumer-groupid sp_sad_asr_group_np_20191028_v3 \ + --session-timeout-ms 60000 \ + --seg-auto-offset-reset smallest \ + --asr-producer-topics asr_topic1_t1 \ + --father-path /data/mfs/k8s/speech_pipeline/raw \ + --hkust-path /home/app/hkust/kaldi/egs/hkust/s5_daihou \ + --num-job 40 +# --is-comp 1 + + +## online +#python2 -u asr_pipline.py --kafka-host daasoffline1.kafka.dc.puhuifinance.com:6667,daasoffline2.kafka.dc.puhuifinance.com:6667,daasoffline3.kafka.dc.puhuifinance.com:6667,daasoffline4.kafka.dc.puhuifinance.com:6667,daasoffline5.kafka.dc.puhuifinance.com:6667 \ +# --seg-consumer-topics sp_sad_topic \ +# --seg-consumer-groupid sp_sad_asr_group \ +# --session-timeout-ms 60000 \ +# --seg-auto-offset-reset largest \ +# --asr-producer-topics sp_asr_topic \ +# --father-path /data/mfs/k8s/speech_pipeline/sad \ +# --hkust-path /home/app/asr_pipline/kaldi/egs/hkust/s5_daihou \ +# --num-job 10 diff --git a/egs/aishell/asr1/local/aishell_data_prep_infer.sh b/egs/aishell/asr1/local/aishell_data_prep_infer.sh index c598ed5fc13..157b5c8dd4c 100755 --- a/egs/aishell/asr1/local/aishell_data_prep_infer.sh +++ b/egs/aishell/asr1/local/aishell_data_prep_infer.sh @@ -12,7 +12,7 @@ if [ $# != 2 ]; then fi aishell_audio_dir=$1 -aishell_text=$2/aishell_transcript_v0.8.txt.0604.relabeled.0923.v2 +aishell_text=$2/aishell_transcript.txt #train_dir=data/local/train #dev_dir=data/local/dev From 5e8c7bf0f345836f6bf549f01fd0821aa41cc7c6 Mon Sep 17 00:00:00 2001 From: nipeng Date: Mon, 28 Oct 2019 19:05:38 +0800 Subject: [PATCH 13/23] =?UTF-8?q?=E4=BF=AE=E6=94=B9=E8=BE=93=E5=87=BA?= =?UTF-8?q?=E6=96=87=E4=BB=B6=E8=B7=AF=E5=BE=84?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- egs/aishell/asr1/infer.sh | 23 +++++++++++++------ .../asr1/local/aishell_data_prep_infer.sh | 17 ++++++++------ 2 files changed, 26 insertions(+), 14 deletions(-) diff --git a/egs/aishell/asr1/infer.sh b/egs/aishell/asr1/infer.sh index 731d21832d0..974eb830d31 100755 --- a/egs/aishell/asr1/infer.sh +++ b/egs/aishell/asr1/infer.sh @@ -6,6 +6,15 @@ . ./path.sh || exit 1; . ./cmd.sh || exit 1; +if [ $# != 2 ]; then + echo "Usage: $0 " + exit 1; +else + data=$1 # 输入文件夹。 + flag=$2 # 唯一标识，防止文件覆盖。 +fi + + # general configuration backend=pytorch stage=0 # start from 0 if you need to start from data preparation @@ -37,7 +46,7 @@ n_average=10 #data_url=www.openslr.org/resources/33 # data=/data/nipeng/2019-10-07-aishell -data=/data/nipeng/2019-1001-1020-aishell +# data=/data/nipeng/2019-1001-1020-aishell # exp tag tag="" # tag for managing experiments. @@ -53,7 +62,7 @@ set -o pipefail train_set=train_sp train_dev=dev #recog_set="dev test" -recog_set="infer" +recog_set=infer_${flag} # if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then # echo "stage -1: Data Download" @@ -65,9 +74,9 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ### Task dependent. You have to make data the following preparation part by yourself. ### But you can utilize Kaldi recipes in most cases echo "stage 0: Data preparation" - local/aishell_data_prep_infer.sh ${data}/data_aishell/wav ${data}/data_aishell/transcript + local/aishell_data_prep_infer.sh ${data}/data_aishell/wav ${data}/data_aishell/transcript ${flag} # remove space in text - for x in infer; do + for x in ${recog_set}; do cp data/${x}/text data/${x}/text.org paste -d " " <(cut -f 1 -d" " data/${x}/text.org) <(cut -f 2- -d" " data/${x}/text.org | tr -d " ") \ > data/${x}/text @@ -81,7 +90,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then ### Task dependent. You have to design training and dev sets by yourself. ### But you can utilize Kaldi recipes in most cases echo "stage 1: Feature Generation" - fbankdir=fbank + fbankdir=fbank_${flag} # Generate the fbank features; by default 80-dimensional fbanks with pitch on each frame # steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 32 --write_utt2num_frames true \ # data/train exp/make_fbank/train ${fbankdir} @@ -90,8 +99,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then # data/dev exp/make_fbank/dev ${fbankdir} # utils/fix_data_dir.sh data/dev steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 10 --write_utt2num_frames true \ - data/infer exp/make_fbank/infer ${fbankdir} - utils/fix_data_dir.sh data/infer + data/${recog_set} exp/make_fbank/${recog_set} ${fbankdir} + utils/fix_data_dir.sh data/${recog_set} # # speed-perturbed # utils/perturb_data_dir_speed.sh 0.9 data/train data/temp1 diff --git a/egs/aishell/asr1/local/aishell_data_prep_infer.sh b/egs/aishell/asr1/local/aishell_data_prep_infer.sh index f4b75d102a7..7d5ec817521 100755 --- a/egs/aishell/asr1/local/aishell_data_prep_infer.sh +++ b/egs/aishell/asr1/local/aishell_data_prep_infer.sh @@ -5,24 +5,27 @@ . ./path.sh || exit 1; -if [ $# != 2 ]; then - echo "Usage: $0 " - echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript" +if [ $# != 3 ]; then + echo "Usage: $0 " + echo " $0 /export/a05/xna/data/data_aishell/wav /export/a05/xna/data/data_aishell/transcript uuid" exit 1; fi aishell_audio_dir=$1 aishell_text=$2/aishell_transcript_v0.8.txt.0604.relabeled.0923.v2 +flag=$3 #train_dir=data/local/train #dev_dir=data/local/dev #test_dir=data/local/test -infer_dir=data/local/infer -tmp_dir=data/local/tmp +raw_infer_dir=data/infer_${flag} +infer_dir=data/local/infer_${flag} +tmp_dir=data/local/tmp_${flag} #mkdir -p $train_dir #mkdir -p $dev_dir #mkdir -p $test_dir +mkdir -p $raw_infer_dir mkdir -p $infer_dir mkdir -p $tmp_dir @@ -56,10 +59,10 @@ for dir in $infer_dir; do utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt done -mkdir -p data/infer +## mkdir -p data/infer_${flag} for f in spk2utt utt2spk wav.scp text; do - cp $infer_dir/$f data/infer/$f || exit 1; + cp $infer_dir/$f ${raw_infer_dir}/$f || exit 1; done echo "$0: AISHELL data preparation succeeded" From aaef5d15f0253ebd5ca7bc804e9174decf9f8a6c Mon Sep 17 00:00:00 2001 From: root Date: Tue, 29 Oct 2019 09:42:53 +0800 Subject: [PATCH 14/23] =?UTF-8?q?=E5=A2=9E=E5=8A=A0conda=E6=BA=90?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- egs/aishell/asr1/asr_pipline.py | 5 ++++- tools/Makefile | 8 ++++++++ 2 files changed, 12 insertions(+), 1 deletion(-) diff --git a/egs/aishell/asr1/asr_pipline.py b/egs/aishell/asr1/asr_pipline.py index 50bb04b88e4..a8f9b24b6fb 100644 --- a/egs/aishell/asr1/asr_pipline.py +++ b/egs/aishell/asr1/asr_pipline.py @@ -388,7 +388,10 @@ def _asr_cmd(self, wav_path): :param wav_path: 需识别音频存放目录 :return: """ - decode_cmd = "./infer.sh {}".format(wav_path) + print("==1", wav_path) + flag = wav_path.split("_")[-2] + print(flag) + decode_cmd = "./infer.sh {} {}".format(wav_path, flag) print(decode_cmd) os.system(decode_cmd) ## return merge_dict diff --git a/tools/Makefile b/tools/Makefile index d3c6a38ede3..11f9cd51f01 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -68,14 +68,22 @@ venv: miniconda.sh . venv/bin/activate && conda update -y conda . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ + . venv/bin/activate && conda config --add channels https://mirror.tuna.tsinghua.edu.cn/anaconda/pkgs/main/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/bioconda/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/menpo/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y python=$(PYTHON_VERSION) . venv/bin/activate && conda info -a espnet.done: venv . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ + . venv/bin/activate && conda config --add channels https://mirror.tuna.tsinghua.edu.cn/anaconda/pkgs/main/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/bioconda/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/menpo/ + . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y $(CONDA_PYTORCH) -c pytorch . venv/bin/activate && pip install -e .. From 9cb916b6b4da826c59e6cb1e69cbd5748eabaa42 Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 09:57:10 +0800 Subject: [PATCH 15/23] =?UTF-8?q?=E5=AE=89=E8=A3=85pytorch=E4=BD=BF?= =?UTF-8?q?=E7=94=A8=E6=8C=87=E5=AE=9A=E7=9A=84=E6=B8=85=E5=8D=8E=E6=BA=90?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- tools/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/Makefile b/tools/Makefile index 11f9cd51f01..436085278a6 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -85,7 +85,7 @@ espnet.done: venv . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/menpo/ . venv/bin/activate && conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ . venv/bin/activate && conda config --set show_channel_urls yes - . venv/bin/activate && conda install -y $(CONDA_PYTORCH) -c pytorch + . venv/bin/activate && conda install -y $(CONDA_PYTORCH) pytorch . venv/bin/activate && pip install -e .. touch espnet.done endif From 9e0d678dcec11f1497104ccc274af4ab466af62a Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 10:17:24 +0800 Subject: [PATCH 16/23] =?UTF-8?q?=E5=88=A0=E9=99=A4cudatookit?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- tools/Makefile | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/tools/Makefile b/tools/Makefile index 436085278a6..43e29c3b995 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -20,13 +20,13 @@ CONDA_URL := https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64. GCC_VERSION := $(shell gcc -dumpversion) -ifneq ($(shell which nvidia-smi),) # 'nvcc' found -CONDA_PYTORCH := pytorch=$(TH_VERSION) cudatoolkit=$(CUDA_VERSION) -CUDA_DEPS := cupy.done -else +# ifneq ($(shell which nvidia-smi),) # 'nvcc' found +# CONDA_PYTORCH := pytorch=$(TH_VERSION) cudatoolkit=$(CUDA_VERSION) +# CUDA_DEPS := cupy.done +# else CONDA_PYTORCH := pytorch-cpu=$(TH_VERSION) CUDA_DEPS := -endif +# endif .PHONY: all clean From 625f19e81eb8f6abfe8df9298915e6b0cb047c0a Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 15:02:26 +0800 Subject: [PATCH 17/23] add --- egs/aishell/asr1/asr_pipline.py | 148 ++++++++++++-------------------- egs/aishell/asr1/asr_pipline.sh | 4 +- egs/aishell/asr1/infer.sh | 7 +- 3 files changed, 63 insertions(+), 96 deletions(-) mode change 100644 => 100755 egs/aishell/asr1/asr_pipline.sh diff --git a/egs/aishell/asr1/asr_pipline.py b/egs/aishell/asr1/asr_pipline.py index 50bb04b88e4..b88f66eaf64 100644 --- a/egs/aishell/asr1/asr_pipline.py +++ b/egs/aishell/asr1/asr_pipline.py @@ -28,7 +28,6 @@ def norm_aishell_data(indir): - # /tmp/asr_9hmlx3zl outdir = indir + "_norm" for path in os.listdir(indir): fname = path.split(".")[0] @@ -52,10 +51,22 @@ def gen_sp_wav_and_get_path_one(wav_temppath, audio_id, sound, item, k): if sp_wav.frame_rate != 8000: sp_wav = sp_wav.set_frame_rate(8000) sp_wav.export(save_cut_path, format="wav") - print("==3", save_cut_path) return save_cut_path +def load_hyp_file(path): + merge_dict = {} + with codecs.open(path) as f: + for line in f: + l = line.split("(") + if len(l) == 2: + _text, _id = l + text = _text.strip().replace(" ", "") + path = _id.split("-")[0] + ".wav" + merge_dict[path] = text + return merge_dict + + def get_parser(): parser = argparse.ArgumentParser(description='语音识别主函数参数') parser.add_argument("-kh", "--kafka-host", @@ -101,7 +112,7 @@ def get_parser(): help="multiprocess number of workers") parser.add_argument("-cg", "--consumer-gap", - type=int, default=2, + type=int, default=10, help="kafka consumer msg num") parser.add_argument("-ptm", "--poll-timeout-ms", @@ -146,6 +157,7 @@ def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, compression_type="gzip", max_request_size=1024 * 1024 * 20) self.asr_producer_topics = asr_producer_topics # ASR生产者topic + self.redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") self.is_comp = is_comp if is_comp: @@ -175,7 +187,6 @@ def asr_pipline_from_kafka(self, father_path): :param father_path: 切分来源音频存放父目录 :return: """ - # redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") while True: if not self.from_client: self._get_from_client() @@ -185,9 +196,7 @@ def asr_pipline_from_kafka(self, father_path): msgs = [] for tp, _msgs in tp_msgs.items(): msgs.extend(_msgs) - print(len(msgs)) self.batch_asr_pipline(father_path, msgs) - break def batch_asr_pipline(self, father_path, msgs): """ @@ -207,41 +216,35 @@ def batch_asr_pipline(self, father_path, msgs): batch_voice_data = {} batch_merge_dict = None -# try: - redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") - for msg in msgs: - if msg is not None: - audio_id = json.loads(msg.value).get('audio_id', '') - print("==2", audio_id) - if not redis_client.get('asr_espnet_' + str(audio_id)): - # step1: 从kafka消息中提取value信息 - voice_data = json.loads(msg.value) - voice_data['start_asr'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) - batch_voice_data[audio_id] = voice_data # 后面继续往 voice_data 追加数据 - # step2：根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav - batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data) - # step3: 语音识别, 并获取merge的text - # batch results - print("==4", wav_temppath) - wav_normpath = norm_aishell_data(wav_temppath) - merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} - #batch_merge_dict = self.split_merge_dict(merge_dict) -# except Exception as e: -# print("bebefore commit to kafka error log: %s, msg: %s" % (e, "")) -# finally: -# # step7: 删除临时音频的文件夹和语音识别结果的临时文件 -# #shutil.rmtree(wav_temppath) -# pass - return + try: + for msg in msgs: + if msg is not None: + audio_id = json.loads(msg.value).get('audio_id', '') + if not redis_client.get('asr_espnet_' + str(audio_id)): + # step1: 从kafka消息中提取value信息 + voice_data = json.loads(msg.value) + voice_data['start_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + batch_voice_data[audio_id] = voice_data # 后面继续往 voice_data 追加数据 + # step2：根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav + batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data) + # step3: 语音识别, 并获取merge的text + # batch results + wav_normpath = norm_aishell_data(wav_temppath) + merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} + batch_merge_dict = self.split_merge_dict(merge_dict) + except Exception as e: + print("bebefore commit to kafka error log: %s, msg: %s" % (e, "")) + finally: + # step7: 删除临时音频的文件夹和语音识别结果的临时文件 + shutil.rmtree(wav_temppath) + shutil.rmtree(wav_normpath) for audio_id, voice_data in batch_voice_data.items(): # step4: 写入kafka新的topic try: - voice_data["merge_dict"] = batch_merge_dict[audio_id] - voice_data['end_asr'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + voice_data["merge_dict_espnet"] = batch_merge_dict[audio_id] + voice_data['end_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) flag = self._kafka_producers(voice_data) if voice_data.get("merge_dict", {}) else False - voice_data['step4_end_time'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) - print("asr_output:", json.dumps(voice_data, ensure_ascii=True)) except Exception as e: print("before commit to kafka error log: %s, msg: %s" % (e, "")) try: @@ -249,7 +252,7 @@ def batch_asr_pipline(self, father_path, msgs): except CommitFailedError: print('commit error!') for audio_id, _ in batch_voice_data.items(): - redis_client.set('asr_espnet_' + str(audio_id), 1, ex=24 * 3600) + self.redis_client.set('asr_espnet_' + str(audio_id), 1, ex=24 * 3600) return flag @@ -300,57 +303,6 @@ def filter_conditions(self, voice_data): raise RuntimeError('This data segments is null !') self.comp_num += 1 - def _copy_wav(self, father_path, voice_data, temp_path): - """ - step2: 根据msg信息从切分好的数据源复制到目标位置 - :param father_path: 切分来源音频存放父目录 - :param voice_data: kafka来源信息中的包含切分音频的信息 - :param temp_path: 音频复制目标目录 - :return: wav_lst: - """ - source = voice_data.get("source", "") - date = voice_data.get("date", "") - audio_id = voice_data.get("audio_id", "") - - value_dict = voice_data.get("segments", {}) - wav_lst = [] # 音频list - - for chn, split_time in value_dict.items(): - for si in split_time: - split_path = "%s_%s_%s_%s.wav" % (audio_id, si.get("start"), si.get("end"), chn) - wav_path = os.path.join(father_path, source, date, audio_id, split_path) - shutil.copy(wav_path, temp_path) - wav_lst.append(wav_path) - return wav_lst - - def gen_sp_wav_and_get_path(self, father_path, wav_temppath, voice_data): - """ - step2: 根据msg信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav - :param father_path: 音频来源目录,结果存储 - :param wav_temppath: kafka来源信息中的包含切分音频的信息 - :return: wav_lst: - """ - wav_lst = [] # wav 音频存储 - source = voice_data["source"] - code = None if source != 'infobird' else 'pcm_s16le' - date = voice_data["date"] - audio_id = voice_data["audio_id"] - wav_father_path = os.path.join(father_path, source, date) # /data/mfs/k8s/speech_pipeline/raw/{source}/{date} - for k, v in voice_data['segments'].items(): - if k in ["gk", "kf"] and len(v) > 0: - wav_name = "%s_%s.wav" % (audio_id, k) - raw_wav_path = os.path.join(wav_father_path, wav_name) - sound = AudioSegment.from_file(raw_wav_path, codec=code, format="wav") - for item in v: - cut_wav_name = "%s_%s_%s_%s.wav" % (audio_id, item['start'], item['end'], k) - save_cut_path = os.path.join(wav_temppath, cut_wav_name) - sp_wav = sound[int(item['start']):int(item['end'])] - if sp_wav.frame_rate != 8000: - sp_wav = sp_wav.set_frame_rate(8000) - sp_wav.export(save_cut_path, format="wav") - wav_lst.append(save_cut_path) - return wav_lst - def gen_sp_wav_and_get_path_mp(self, father_path, wav_temppath, batch_voice_data): """ multiprocess generator @@ -388,10 +340,24 @@ def _asr_cmd(self, wav_path): :param wav_path: 需识别音频存放目录 :return: """ - decode_cmd = "./infer.sh {}".format(wav_path) - print(decode_cmd) + flag = wav_path.split("/")[-1].replace("asr", "").replace("norm", "") + decode_cmd = "./infer.sh {} {} {}".format(wav_path, flag, self.num_job) os.system(decode_cmd) - ## return merge_dict + decode_path = "exp/train_sp_pytorch_train/decode_infer_%s_decode_lm/hyp.trn" %flag + merge_dict = load_hyp_file(decode_path) + + ## 删除临时目录 + data_dir = "data/infer_%s" %flag + fbank_dir = "fbank_%s" %flag + dump_dir = "dump/infer_%s" %flag + decode_dir = "exp/train_sp_pytorch_train/decode_infer_%s_decode_lm" %flag + dump_feats_dir = "exp/dump_feats/recog/infer_%s" %flag + make_fbank_dir = "exp/make_fbank/infer_%s" %flag + + for _dir in [data_dir, fbank_dir, dump_dir, dump_feats_dir, make_fbank_dir, decode_dir]: + os.system("rm -rf %s" %_dir) + + return merge_dict def _kafka_producers(self, voice_data): """ diff --git a/egs/aishell/asr1/asr_pipline.sh b/egs/aishell/asr1/asr_pipline.sh old mode 100644 new mode 100755 index 52af73ac321..157d38245ba --- a/egs/aishell/asr1/asr_pipline.sh +++ b/egs/aishell/asr1/asr_pipline.sh @@ -13,12 +13,12 @@ python3 asr_pipline.py \ --kafka-host daasoffline1.kafka.dc.puhuifinance.com:6667,daasoffline2.kafka.dc.puhuifinance.com:6667,daasoffline3.kafka.dc.puhuifinance.com:6667,daasoffline4.kafka.dc.puhuifinance.com:6667,daasoffline5.kafka.dc.puhuifinance.com:6667 \ --seg-consumer-topics sp_sad_topic \ --seg-consumer-groupid sp_sad_asr_group_np_20191028_v3 \ - --session-timeout-ms 60000 \ + --session-timeout-ms 30000 \ --seg-auto-offset-reset smallest \ --asr-producer-topics asr_topic1_t1 \ --father-path /data/mfs/k8s/speech_pipeline/raw \ --hkust-path /home/app/hkust/kaldi/egs/hkust/s5_daihou \ - --num-job 40 + --num-job 10 # --is-comp 1 diff --git a/egs/aishell/asr1/infer.sh b/egs/aishell/asr1/infer.sh index 974eb830d31..3274b4b92df 100755 --- a/egs/aishell/asr1/infer.sh +++ b/egs/aishell/asr1/infer.sh @@ -6,12 +6,13 @@ . ./path.sh || exit 1; . ./cmd.sh || exit 1; -if [ $# != 2 ]; then - echo "Usage: $0 " +if [ $# != 3 ]; then + echo "Usage: $0 " exit 1; else data=$1 # 输入文件夹。 flag=$2 # 唯一标识，防止文件覆盖。 + nj_decoder=$3 # decode workers fi @@ -223,7 +224,7 @@ expdir=exp/${expname} if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then echo "stage 5: Decoding" - nj=40 + nj=${nj_decoder} if [[ $(get_yaml.py ${train_config} model-module) = *transformer* ]]; then recog_model=model.last${n_average}.avg.best # average_checkpoints.py --backend ${backend} \ From e7fe3a168a002c082cc754f032c0eb76ba63c12c Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 16:20:56 +0800 Subject: [PATCH 18/23] =?UTF-8?q?add:=20=E8=BF=87=E6=BB=A4=E5=85=B3?= =?UTF-8?q?=E9=94=AE=E8=AF=8D?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- egs/aishell/asr1/asr_pipline.py | 81 ++++++++++++++++++++++----------- egs/aishell/asr1/asr_pipline.sh | 2 +- 2 files changed, 55 insertions(+), 28 deletions(-) diff --git a/egs/aishell/asr1/asr_pipline.py b/egs/aishell/asr1/asr_pipline.py index b88f66eaf64..69abba444d5 100644 --- a/egs/aishell/asr1/asr_pipline.py +++ b/egs/aishell/asr1/asr_pipline.py @@ -7,6 +7,7 @@ sys.setdefaultencoding("utf-8") import os +import re import time import json import redis @@ -27,8 +28,8 @@ from elasticsearch import Elasticsearch, helpers -def norm_aishell_data(indir): - outdir = indir + "_norm" +def norm_aishell_data(indir, outdir): + os.system("mkdir -p %s" %outdir) for path in os.listdir(indir): fname = path.split(".")[0] os.system("mkdir -p %s/data_aishell/wav/infer/%s" %(outdir, fname)) @@ -41,7 +42,6 @@ def norm_aishell_data(indir): for path in os.listdir(indir): fname = path.split(".")[0] fout.write("%s %s\n" %(fname, "哈哈")) - return outdir def gen_sp_wav_and_get_path_one(wav_temppath, audio_id, sound, item, k): @@ -104,7 +104,7 @@ def get_parser(): help="hkust的绝对目录,即kaldi的hkust目录") parser.add_argument("-nj", "--num-job", - type=int, default=10, + type=int, default=2, help="hkust num job default 10") parser.add_argument("-nw", "--num-workers", @@ -124,6 +124,22 @@ def get_parser(): return args +class WarnKeyword(object): + def __init__(self): + kws = ["狗杂种", "操你妈", "傻逼", "他妈的","你妈逼", + "狗日的","王八蛋", "妈了个逼","婊子", "去你妈", + "我操", "我草","贱人", "被车撞死", "搞死", + "密码给我", "老赖","曝通讯录", "所有联系人", "不要脸", + "去死","要不要脸", "打爆你", + ] + self.p_kw = re.compile("|".join(kws)) + + def process(self, text): + l = self.p_kw.search(text) + rst = True if l else False + return rst + + class ASR(object): def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, session_timeout_ms=60000, seg_auto_offset_reset="largest", @@ -158,6 +174,7 @@ def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, max_request_size=1024 * 1024 * 20) self.asr_producer_topics = asr_producer_topics # ASR生产者topic self.redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") + self.kw_client = WarnKeyword() self.is_comp = is_comp if is_comp: @@ -181,6 +198,17 @@ def _get_from_client(self): auto_offset_reset=self.seg_auto_offset_reset) # 消费重置偏移量 self.from_client.subscribe(self.seg_consumer_topics) # 切分的消费者topic + def filter_msg_keyword(self, msgs): + rst = [] + for msg in msgs: + data = json.loads(msg.value) + merge_dict = msg.get("merge_dict") + if merge_dict: + text = ";".join(merge_dict.values()) + if self.kw_client.process(text): + rst.append(msg) + return rst + def asr_pipline_from_kafka(self, father_path): """ 获取kafka的数据流，并进行识别，合并，标点，存入es @@ -196,22 +224,19 @@ def asr_pipline_from_kafka(self, father_path): msgs = [] for tp, _msgs in tp_msgs.items(): msgs.extend(_msgs) - self.batch_asr_pipline(father_path, msgs) + + ## 从kaldi识别结果中，过滤出包含告警关键词的数据 + msgs = self.filter_msg_keyword(msgs) + + if msgs: + self.batch_asr_pipline(father_path, msgs) + def batch_asr_pipline(self, father_path, msgs): - """ - 单个kafka消息消费: - 1)从kafka消息中提取value信息; - 2)根据msg.value信息从切分好的数据源复制到目标位置; - 3)进行语音识别,放进merge_dict里; - 4)写入kafka新的topic; - 7)删除临时音频的文件夹和语音识别结果的临时文件. - :param father_path: 切分来源音频存放父目录; - :param msg: 从kafka获取的完整消息; - :return: - """ flag = False # flag 是否语音识别成功并存入kafka wav_temppath = tempfile.mkdtemp(prefix="asr_") + wav_normpath = wav_temppath + "_norm" + batch_wav_lst = [] batch_voice_data = {} batch_merge_dict = None @@ -220,33 +245,35 @@ def batch_asr_pipline(self, father_path, msgs): for msg in msgs: if msg is not None: audio_id = json.loads(msg.value).get('audio_id', '') - if not redis_client.get('asr_espnet_' + str(audio_id)): - # step1: 从kafka消息中提取value信息 + if not self.redis_client.get('asr_espnet_' + str(audio_id)): voice_data = json.loads(msg.value) voice_data['start_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) batch_voice_data[audio_id] = voice_data # 后面继续往 voice_data 追加数据 - # step2：根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav + + # 根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data) - # step3: 语音识别, 并获取merge的text - # batch results - wav_normpath = norm_aishell_data(wav_temppath) + # 语音识别, 并获取merge的text + # 整理成espnet需要的数据格式 + norm_aishell_data(wav_temppath, wav_normpath) + # espnet decode merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} + # merge batch batch_merge_dict = self.split_merge_dict(merge_dict) except Exception as e: - print("bebefore commit to kafka error log: %s, msg: %s" % (e, "")) + print("asr cmd error log: %s, msg: %s" % (e, "")) finally: - # step7: 删除临时音频的文件夹和语音识别结果的临时文件 + # 删除临时音频的文件夹和语音识别结果的临时文件 shutil.rmtree(wav_temppath) shutil.rmtree(wav_normpath) for audio_id, voice_data in batch_voice_data.items(): - # step4: 写入kafka新的topic + # 写入kafka新的topic try: voice_data["merge_dict_espnet"] = batch_merge_dict[audio_id] voice_data['end_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) - flag = self._kafka_producers(voice_data) if voice_data.get("merge_dict", {}) else False + flag = self._kafka_producers(voice_data) if voice_data.get("merge_dict_espnet", {}) else False except Exception as e: - print("before commit to kafka error log: %s, msg: %s" % (e, "")) + print("kafka producer error log: %s, msg: %s" % (e, "")) try: self.from_client.commit() except CommitFailedError: diff --git a/egs/aishell/asr1/asr_pipline.sh b/egs/aishell/asr1/asr_pipline.sh index 157d38245ba..814763c2911 100755 --- a/egs/aishell/asr1/asr_pipline.sh +++ b/egs/aishell/asr1/asr_pipline.sh @@ -11,7 +11,7 @@ # 测试需要在代码内修改生产则kafka host python3 asr_pipline.py \ --kafka-host daasoffline1.kafka.dc.puhuifinance.com:6667,daasoffline2.kafka.dc.puhuifinance.com:6667,daasoffline3.kafka.dc.puhuifinance.com:6667,daasoffline4.kafka.dc.puhuifinance.com:6667,daasoffline5.kafka.dc.puhuifinance.com:6667 \ - --seg-consumer-topics sp_sad_topic \ + --seg-consumer-topics sp_asr_topic \ --seg-consumer-groupid sp_sad_asr_group_np_20191028_v3 \ --session-timeout-ms 30000 \ --seg-auto-offset-reset smallest \ From 1bc353d92beae8fdaf3c9920381bf56c8cf80381 Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 17:30:22 +0800 Subject: [PATCH 19/23] add: pip install --- tools/Makefile | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tools/Makefile b/tools/Makefile index 43e29c3b995..33c991e07aa 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -59,6 +59,7 @@ espnet.done: venv . venv/bin/activate; pip install -e .. . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ torch==$(TH_VERSION) + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka elasticsearch touch espnet.done else miniconda.sh: @@ -87,6 +88,7 @@ espnet.done: venv . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y $(CONDA_PYTORCH) pytorch . venv/bin/activate && pip install -e .. + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka elasticsearch touch espnet.done endif From b836d488336a5cf47fd819f8ccfe8849d707a9b3 Mon Sep 17 00:00:00 2001 From: nipeng Date: Tue, 29 Oct 2019 17:47:26 +0800 Subject: [PATCH 20/23] add: pip index --- tools/Makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tools/Makefile b/tools/Makefile index 33c991e07aa..e1e8cda612b 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -128,7 +128,7 @@ warp-transducer.done: espnet.done chainer_ctc.done: espnet.done rm -rf chainer_ctc git clone https://github.com/jheymann85/chainer_ctc.git - . venv/bin/activate; pip install cython + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ cython . venv/bin/activate; cd chainer_ctc && chmod +x install_warp-ctc.sh && ./install_warp-ctc.sh ; true . venv/bin/activate; cd chainer_ctc && pip install . touch chainer_ctc.done @@ -147,7 +147,7 @@ mecab.done: espnet.done cd mecab/mecab-ipadic && ./configure --with-charset=utf8 --with-mecab-config=$(PWD)/mecab/bin/mecab-config --prefix=$(PWD)/mecab && $(MAKE) && $(MAKE) install cd mecab && git clone --depth 1 https://github.com/neologd/mecab-ipadic-neologd.git cd mecab/mecab-ipadic-neologd && export PATH=$(PWD)/mecab/bin:$(PATH) && ./bin/install-mecab-ipadic-neologd -n -y -p $(PWD)/mecab/mecab-ipadic-neologd - . venv/bin/activate; pip install mojimoji pykakasi + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ mojimoji pykakasi . venv/bin/activate; if [ `python --version | cut -c 8` -eq 3 ]; then \ if [ ! -e swig.done ]; then \ rm -rf swig; \ From 116b79585c0b76ff0062d181462ace1072295736 Mon Sep 17 00:00:00 2001 From: nipeng Date: Wed, 30 Oct 2019 10:14:48 +0800 Subject: [PATCH 21/23] =?UTF-8?q?=E5=A2=9E=E5=8A=A0=E6=98=AF=E5=90=A6?= =?UTF-8?q?=E5=91=BD=E4=B8=AD=E5=91=8A=E8=AD=A6=E5=85=B3=E9=94=AE=E8=AF=8D?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- docker/infer.sh | 12 +- .../asr1/{asr_pipline.py => asr_pipeline.py} | 133 ++++++------------ .../asr1/{asr_pipline.sh => asr_pipeline.sh} | 5 +- egs/aishell/asr1/test.py | 11 +- tools/Makefile | 2 +- 5 files changed, 58 insertions(+), 105 deletions(-) rename egs/aishell/asr1/{asr_pipline.py => asr_pipeline.py} (76%) rename egs/aishell/asr1/{asr_pipline.sh => asr_pipeline.sh} (96%) mode change 100755 => 100644 diff --git a/docker/infer.sh b/docker/infer.sh index e7b6ec914a1..1e8de03482d 100755 --- a/docker/infer.sh +++ b/docker/infer.sh @@ -24,11 +24,11 @@ do eval ${ext}=$2 frombreak=false break 2 - fi - done + fi + done done if ${frombreak} ; then - echo "bad option $1" + echo "bad option $1" exit 1 fi ;; @@ -47,7 +47,7 @@ fi from_tag="cpu" if [ ! "${docker_gpu}" == "-1" ]; then if [ -z "${docker_cuda}" ]; then - # If the docker_cuda is not set, the program will automatically + # If the docker_cuda is not set, the program will automatically # search the installed version with default configurations (apt) docker_cuda=$( nvcc -V | grep release ) docker_cuda=${docker_cuda#*"release "} @@ -75,7 +75,7 @@ fi if [ ${docker_user} = true ]; then # Build a container with the user account container_tag="${from_tag}-user-${HOME##*/}" - docker_image=$( docker images -q espnet/espnet:${container_tag} ) + docker_image=$( docker images -q espnet/espnet:${container_tag} ) if ! [[ -n ${docker_image} ]]; then echo "Building docker image..." build_args="--build-arg FROM_TAG=${from_tag}" @@ -130,7 +130,7 @@ if [ ! -z "${docker_env}" ]; then docker_env=$(echo ${docker_env} | tr "," "\n") for i in ${docker_env[@]} do - this_env="-e $i ${this_env}" + this_env="-e $i ${this_env}" done fi diff --git a/egs/aishell/asr1/asr_pipline.py b/egs/aishell/asr1/asr_pipeline.py similarity index 76% rename from egs/aishell/asr1/asr_pipline.py rename to egs/aishell/asr1/asr_pipeline.py index 69abba444d5..408c8043b48 100644 --- a/egs/aishell/asr1/asr_pipline.py +++ b/egs/aishell/asr1/asr_pipeline.py @@ -11,7 +11,6 @@ import time import json import redis -import jieba import codecs import shutil import hashlib @@ -25,7 +24,6 @@ from pydub import AudioSegment from kafka import KafkaConsumer, KafkaProducer from kafka.errors import CommitFailedError -from elasticsearch import Elasticsearch, helpers def norm_aishell_data(indir, outdir): @@ -104,8 +102,8 @@ def get_parser(): help="hkust的绝对目录,即kaldi的hkust目录") parser.add_argument("-nj", "--num-job", - type=int, default=2, - help="hkust num job default 10") + type=int, default=10, + help="espnet decode num job default 10") parser.add_argument("-nw", "--num-workers", type=int, default=18, @@ -119,12 +117,11 @@ def get_parser(): type=int, default=60000, help="") - parser.add_argument('--is-comp', help='1为补数,0为不进入补数筛选逻辑', default=0, type=int) args = parser.parse_args() return args -class WarnKeyword(object): +class DetectAlarmKeyword(object): def __init__(self): kws = ["狗杂种", "操你妈", "傻逼", "他妈的","你妈逼", "狗日的","王八蛋", "妈了个逼","婊子", "去你妈", @@ -135,15 +132,14 @@ def __init__(self): self.p_kw = re.compile("|".join(kws)) def process(self, text): - l = self.p_kw.search(text) - rst = True if l else False + rst = self.p_kw.findall(text) return rst class ASR(object): def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, session_timeout_ms=60000, seg_auto_offset_reset="largest", - asr_producer_topics="asr_topic1", num_job=10, is_comp=0, + asr_producer_topics="asr_topic1", num_job=10, poll_timeout_ms=60000, consumer_gap=None, num_workers=cpu_count()): """ :param kafka_servers: kafka host:port @@ -154,7 +150,6 @@ def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, 源码定义: {'smallest': 'earliest', 'largest': 'latest'} :param asr_producer_topics: 语音是被的生产者topic,默认值:asr_topic1 :param num_job: 语音识别的线程数 - :param is_comp: 是否为补数逻辑,是的话会进入时间判断和es数据判断 """ self.kafka_servers = kafka_servers self.seg_consumer_groupid = seg_consumer_groupid @@ -174,20 +169,7 @@ def __init__(self, kafka_servers, seg_consumer_topics, seg_consumer_groupid, max_request_size=1024 * 1024 * 20) self.asr_producer_topics = asr_producer_topics # ASR生产者topic self.redis_client = redis.Redis(host='192.168.192.202', port=40029, db=0, password="Q8TYmIwQSHNFbLJ2") - self.kw_client = WarnKeyword() - - self.is_comp = is_comp - if is_comp: - self.comp_num = 0 - self.prop = {"HOST1": "192.168.40.37", - "HOST2": "192.168.40.38", - "HOST3": "192.168.40.39", - "PORT": "9200", - "DOC_TYPE": "kf_infobird_call"} - - self.es = Elasticsearch(hosts=[{'host': self.prop['HOST1'], 'port': self.prop['PORT']}, - {'host': self.prop['HOST2'], 'port': self.prop['PORT']}, - {'host': self.prop['HOST3'], 'port': self.prop['PORT']}]) + self.kw_client = DetectAlarmKeyword() def _get_from_client(self): # 消费者切分好的音频kafka消费者 @@ -198,15 +180,13 @@ def _get_from_client(self): auto_offset_reset=self.seg_auto_offset_reset) # 消费重置偏移量 self.from_client.subscribe(self.seg_consumer_topics) # 切分的消费者topic - def filter_msg_keyword(self, msgs): - rst = [] - for msg in msgs: - data = json.loads(msg.value) - merge_dict = msg.get("merge_dict") - if merge_dict: - text = ";".join(merge_dict.values()) - if self.kw_client.process(text): - rst.append(msg) + def check_alarm_keyword(self, merge_dict): + rst = None + if merge_dict: + text = ";".join(merge_dict.values()) + kw_rsp = self.kw_client.process(text) + if kw_rsp: + rst = kw_rsp return rst def asr_pipline_from_kafka(self, father_path): @@ -225,11 +205,7 @@ def asr_pipline_from_kafka(self, father_path): for tp, _msgs in tp_msgs.items(): msgs.extend(_msgs) - ## 从kaldi识别结果中，过滤出包含告警关键词的数据 - msgs = self.filter_msg_keyword(msgs) - - if msgs: - self.batch_asr_pipline(father_path, msgs) + self.batch_asr_pipline(father_path, msgs) def batch_asr_pipline(self, father_path, msgs): @@ -238,7 +214,10 @@ def batch_asr_pipline(self, father_path, msgs): wav_normpath = wav_temppath + "_norm" batch_wav_lst = [] + # 所有数据 batch_voice_data = {} + # 包含报警关键词的数据 + batch_voice_data_imp = {} batch_merge_dict = None try: @@ -247,31 +226,34 @@ def batch_asr_pipline(self, father_path, msgs): audio_id = json.loads(msg.value).get('audio_id', '') if not self.redis_client.get('asr_espnet_' + str(audio_id)): voice_data = json.loads(msg.value) - voice_data['start_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) - batch_voice_data[audio_id] = voice_data # 后面继续往 voice_data 追加数据 - - # 根据msg.value信息获取音频，并按提前提供的音频片段开始结束时间，生成切分后的wav - batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data) - # 语音识别, 并获取merge的text - # 整理成espnet需要的数据格式 - norm_aishell_data(wav_temppath, wav_normpath) - # espnet decode - merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} - # merge batch - batch_merge_dict = self.split_merge_dict(merge_dict) + batch_voice_data[audio_id] = voice_data + if self.check_alarm_keyword(voice_data.get("merge_dict")): + voice_data['start_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + batch_voice_data_imp[audio_id] = voice_data + + if batch_voice_data_imp: + batch_wav_lst = self.gen_sp_wav_and_get_path_mp(father_path, wav_temppath, batch_voice_data_imp) + norm_aishell_data(wav_temppath, wav_normpath) + merge_dict = self._asr_cmd(wav_normpath) if batch_wav_lst else {} + batch_merge_dict = self.split_merge_dict(merge_dict) except Exception as e: - print("asr cmd error log: %s, msg: %s" % (e, "")) + print("asr cmd error log: %s, msg: %s" % (e, "")) finally: # 删除临时音频的文件夹和语音识别结果的临时文件 - shutil.rmtree(wav_temppath) - shutil.rmtree(wav_normpath) + os.system("rm -rf %s" %wav_temppath) + os.system("rm -rf %s" %wav_normpath) for audio_id, voice_data in batch_voice_data.items(): - # 写入kafka新的topic try: - voice_data["merge_dict_espnet"] = batch_merge_dict[audio_id] - voice_data['end_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) - flag = self._kafka_producers(voice_data) if voice_data.get("merge_dict_espnet", {}) else False + if batch_merge_dict and audio_id in batch_merge_dict: + voice_data["merge_dict_espnet"] = batch_merge_dict[audio_id] + voice_data['end_asr_espnet'] = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()) + _alarm_rsp = self.check_alarm_keyword(batch_merge_dict[audio_id]) + if _alarm_rsp: + voice_data["has_alarm_keyword"] = True + voice_data["alarm_keywords"] = json.dumps(_alarm_rsp) + voice_data["asr_model"] = "espnet_20191030" + flag = self._kafka_producers(voice_data) except Exception as e: print("kafka producer error log: %s, msg: %s" % (e, "")) try: @@ -296,40 +278,6 @@ def split_merge_dict(self, merge_dict): split_dict[audio_id][wav_file] = text return split_dict - def filter_conditions(self, voice_data): - """ - 筛选出不需要补数的数据,并抛出异常; - :param voice_data: - :return: - """ - end_sad = voice_data.get("end_sad", "0000-00-00 00:00:00") - audio_id = voice_data.get("audio_id", None) - body = { - "query": {"bool": {"must": [{"term": {"audio_id.keyword": audio_id}}], "must_not": [], "should": []}}, - "from": 0, "size": 10, "sort": [], "aggs": {}} - if end_sad < "2019-07-11 12:00:00" or end_sad > "2019-07-11 21:00:00": - raise RuntimeError('This one\'s end_sad not in time window !') - query = self.es.search(index="asr_text", - doc_type=self.prop['DOC_TYPE'], - scroll='5m', - timeout='30s', - body=body) - results = query['hits']['hits'] # es查询出的结果第一页 - total = query['hits']['total'] # es查询出的结果总量 - print("audio_id:", audio_id, "search num:", total, ",end_sad", end_sad) - if total > 0: - line = results[0]["_source"] - if line.get("cut_text"): - raise RuntimeError('This id already exists and already asr !') - else: - segments = voice_data.get("segments", {}) - for k, v in segments.items(): - if k in ["gk", "kf"] and len(v) > 0: - break - else: - raise RuntimeError('This data segments is null !') - self.comp_num += 1 - def gen_sp_wav_and_get_path_mp(self, father_path, wav_temppath, batch_voice_data): """ multiprocess generator @@ -424,7 +372,6 @@ def _create_id_by_input(self, id=""): seg_consumer_groupid, asr_producer_topics = args.seg_consumer_groupid, args.asr_producer_topics # 语音识别生产者参数 father_path, hkust_path = args.father_path, args.hkust_path # 音频与模型相关参数 num_job = args.num_job # 线程数 - is_comp = args.is_comp # 是否走补数过滤流程 poll_timeout_ms = args.poll_timeout_ms consumer_gap = args.consumer_gap num_workers = args.num_workers @@ -440,7 +387,7 @@ def _create_id_by_input(self, id=""): seg_consumer_topics = [seg_consumer_topics, ] asr = ASR(kafka_servers=kafka_host, seg_consumer_topics=seg_consumer_topics, session_timeout_ms=session_timeout_ms, seg_consumer_groupid=seg_consumer_groupid, seg_auto_offset_reset=seg_auto_offset_reset, - asr_producer_topics=asr_producer_topics, num_job=num_job, is_comp=is_comp, + asr_producer_topics=asr_producer_topics, num_job=num_job, poll_timeout_ms=poll_timeout_ms, consumer_gap=consumer_gap, num_workers=num_workers) asr.asr_pipline_from_kafka(father_path) diff --git a/egs/aishell/asr1/asr_pipline.sh b/egs/aishell/asr1/asr_pipeline.sh old mode 100755 new mode 100644 similarity index 96% rename from egs/aishell/asr1/asr_pipline.sh rename to egs/aishell/asr1/asr_pipeline.sh index 814763c2911..56b85396a9e --- a/egs/aishell/asr1/asr_pipline.sh +++ b/egs/aishell/asr1/asr_pipeline.sh @@ -1,4 +1,4 @@ -#!/usr/bin/bash +#!/bin/bash #vim /etc/hosts ## Entries added by HostAliases. #192.168.176.181 daasoffline1.kafka.dc.puhuifinance.com @@ -9,7 +9,7 @@ # test # 测试需要在代码内修改生产则kafka host -python3 asr_pipline.py \ +/espnet/tools/venv/bin/python asr_pipeline.py \ --kafka-host daasoffline1.kafka.dc.puhuifinance.com:6667,daasoffline2.kafka.dc.puhuifinance.com:6667,daasoffline3.kafka.dc.puhuifinance.com:6667,daasoffline4.kafka.dc.puhuifinance.com:6667,daasoffline5.kafka.dc.puhuifinance.com:6667 \ --seg-consumer-topics sp_asr_topic \ --seg-consumer-groupid sp_sad_asr_group_np_20191028_v3 \ @@ -19,7 +19,6 @@ python3 asr_pipline.py \ --father-path /data/mfs/k8s/speech_pipeline/raw \ --hkust-path /home/app/hkust/kaldi/egs/hkust/s5_daihou \ --num-job 10 -# --is-comp 1 ## online diff --git a/egs/aishell/asr1/test.py b/egs/aishell/asr1/test.py index bfb1cc33981..336c1c7b4af 100644 --- a/egs/aishell/asr1/test.py +++ b/egs/aishell/asr1/test.py @@ -4,7 +4,6 @@ import codecs - def get_infer_result(): path = "/data/nipeng/TTS/espnet/egs/aishell/asr1/exp/train_sp_pytorch_train/decode_infer_decode_lm/hyp.trn" fout = codecs.open("infer.txt", "w") @@ -16,7 +15,15 @@ def get_infer_result(): aname = aid + ".wav" fout.write("%s\t%s\n" %(aname, text)) +def test2(): + from asr_pipeline import DetectAlarmKeyword + kw_client = DetectAlarmKeyword() + text = "你的是傻逼滚按时发大水发老赖法萨芬" + rst = kw_client.process(text) + print(rst) + if __name__ == "__main__": print("ok") - get_infer_result() + #get_infer_result() + test2() diff --git a/tools/Makefile b/tools/Makefile index e1e8cda612b..a58e6dcaf79 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -59,7 +59,7 @@ espnet.done: venv . venv/bin/activate; pip install -e .. . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ torch==$(TH_VERSION) - . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka elasticsearch + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka-python elasticsearch pydub touch espnet.done else miniconda.sh: From 0b6cb891967f16cebe1cd39625cd7917023f3a65 Mon Sep 17 00:00:00 2001 From: nipeng Date: Wed, 30 Oct 2019 11:33:26 +0800 Subject: [PATCH 22/23] add: pydub --- tools/Makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tools/Makefile b/tools/Makefile index a58e6dcaf79..48b31117dba 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -59,7 +59,7 @@ espnet.done: venv . venv/bin/activate; pip install -e .. . venv/bin/activate; pip config set global.index-url https://pypi.mirrors.ustc.edu.cn/simple/ . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ torch==$(TH_VERSION) - . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka-python elasticsearch pydub + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka-python pydub touch espnet.done else miniconda.sh: @@ -88,7 +88,7 @@ espnet.done: venv . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y $(CONDA_PYTORCH) pytorch . venv/bin/activate && pip install -e .. - . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka elasticsearch + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka pydub touch espnet.done endif From 0f8a676ac5762ac9f0a41da53b72b57ed124fb7e Mon Sep 17 00:00:00 2001 From: nipeng Date: Wed, 30 Oct 2019 12:18:58 +0800 Subject: [PATCH 23/23] fix: kafka-python --- tools/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/Makefile b/tools/Makefile index 48b31117dba..8bc37784713 100644 --- a/tools/Makefile +++ b/tools/Makefile @@ -88,7 +88,7 @@ espnet.done: venv . venv/bin/activate && conda config --set show_channel_urls yes . venv/bin/activate && conda install -y $(CONDA_PYTORCH) pytorch . venv/bin/activate && pip install -e .. - . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka pydub + . venv/bin/activate; pip install --index https://pypi.mirrors.ustc.edu.cn/simple/ redis kafka-python pydub touch espnet.done endif