Mandarin Electro-Laryngeal Speech Enhancement Using Cycle-Consistent Generative Adversarial Networks

Electro-laryngeal (EL) speech has poor intelligibility and naturalness, which hampers the popular use of electro-larynx. Voice conversion (VC) can enhance EL speech. However, if to be enhanced is with complicated tone variation rules in Mandarin, enhancement will less effective. This because source (Mandarin speech) target (normal are not strictly parallel. We propose using cycle-consistent generative adversarial networks (CycleGAN, a parallel-free VC framework) continuous Mandarin speech, solve above problem. In proposed framework, generator designed based on neural 2D-Conformer-1D-Transformer-2D-Conformer. Then, we used Mel-Spectrogram instead traditional acoustic features (fundamental frequency, Mel-Cepstrum parameters aperiodicity parameters). At last, converted into waveform signals WaveNet. undertook both subjective objective tests evaluate approach. Compared approaches variable (the average accuracy being 71.59% word error rate 10.85%), our framework increases by 12.12% reduces errors perception 9.15%. towards fixed 29.89% 10.74%), 42.38% 8.59%. Our effectively address problem that The naturalness have been further improved.