Collapsed speech segment detection and suppression for WaveNet vocoder
share
In this paper, we propose a technique to alleviate quality degradation caused by collapsed speech segments sometimes generated by WaveNet vocoder. The effectiveness of WaveNet vocoder to generate natural speech from the acoustic features has been proved in recent works. However, WaveNet vocoder sometimes generates very noisy speech suffering from the collapsed speech segments when only the limited amount of training data is available or significant acoustic mismatches exist between training and testing data. Such corpus limitation and limited model ability easily exist in some speech generation applications, such as voice conversion or speech enhancement. To address this issue, we propose a technique to automatically detect the collapsed speech segments. Moreover, to refine the detected segments, we also propose a waveform generation technique for WaveNet using a linear predictive coding constraint. Verification and subjective tests are conducted to investigate the effectiveness of the proposed techniques. The verification results indicate that the detection technique can detect most collapsed segments. The subjective evaluations in voice conversion demonstrate the generation technique achieves significant speech quality improvement while keeping the same speaker similarity.
READ FULL TEXT