This paper presents a noise-robust voice conversion method with high-quefrency boosting via sub-band cepstrum conversion and fusion based on the bidirectional long short-term memory (BLSTM) neural networks that can convert parameters of vocal tracks of a source speaker into those of a target speaker. With the implementation of state-of-the-art machine learning methods, voice conversion has achieved good performance given abundant clean training data. However, the quality and similarity of the converted voice are significantly degraded compared to that of a natural target voice due to various factors, such as limited training data and noisy input speech from the source speaker. To address the problem of noisy input speech, an architecture of voice conversion with statistical filtering and sub-band cepstrum conversion and fusion is introduced. The impact of noises on the converted voice is reduced by the accurate reconstruction of the sub-band cepstrum and the subsequent statistical filtering. By normalizing the mean and variance of the converted cepstrum to those of the target cepstrum in the training phase, a cepstrum filter was constructed to further improve the quality of the converted voice. The experimental results showed that the proposed method significantly improved the naturalness and similarity of the converted voice compared to the baselines, even with the noisy inputs of source speakers.

In recent years, neural network-based voice conversion methods have been rapidly developed, and many different models and neural networks have been applied in parallel voice conversion. However, the over-smoothing of parametric methods [e.g., bidirectional long short-term memory (BLSTM)] and the waveform collapse of neural vocoders (e.g., WaveNet) still have negative impacts on the quality of the converted voices. To overcome this problem, we propose a BLSTM and WaveNet-based voice conversion method cooperated with waveform collapse suppression by post-processing. This method firstly uses BLSTM to convert the acoustic features between parallel speakers, and then synthesizes pre-converted voice with WaveNet. Subsequently, several alternative iterations of BLSTM post-processing is performed, and the final converted voice is generated by WaveNet. The proposed method can directly generate converted audio waveforms and avoid the waveform-collapsed speech caused by a single WaveNet generation effectively. The experimental results indicate that acoustic features trained by using the BLSTM network could achieve better results than conventional baselines. From our experiments on VCC2018, the usage of WaveNet could alleviate the problem of over-smoothing, which contributes to improving the similarity and naturalness of the final results of voice conversion.