Blind Bandwidth Extension of Speech

Abstract

Up to today telephone speech lacks of perceptual quality and in- telligibility due to bandwidth removal and quantisation artefacts in the encoding process. Blind bandwidth extension artificially regenerates this missing frequency content and thus improves the perceptual quality and intelligibility. This work proposes novel approaches for blind bandwidth extension based on generative deep neural networks with convo- lutional or recurrent architectures. Motivated by the source-filter model of the human speech production, two of the proposed systems decompose the speech signal into spectral envelope and excitation signal. Each are bandwidth-extended with ded- icated networks. First the envelope is extrapolated and then the network extrapolating the excitation signal is trained such that there is no mismatch between the excitation signal and the envelope. By this, we achieve better perceptual quality with lower computational complexity compared to a system that ex- tends the bandwidth of the undecomposed speech signal with a single network. Input and output of all proposed systems is raw time-domain speech without any preprocessing. They are trained with a mixture of adversarial and perceptual loss which further improves the perceptual quality. To avoid mode collapse and a more stable adversarial training, spectral normalisation is employed in the discriminator. The presented systems are compared to previously published systems by objective measures and subjectively by a listening test. An estimate of the computational complexity is given and compared to state of the art speech coding technologies. Objec- tive and subjective tests show that the proposed systems deliver substantial better quality than prior techniques. It was further shown that our systems reduces the word error rate of a speech recognition system. Show more