Pretraining Techniques for Sequence-to-Sequence Voice Conversion
This addresses practical issues in voice conversion for speech synthesis applications, but it is incremental as it adapts existing pretraining techniques to a specific domain.
The paper tackles the problem of unstable training and mispronunciation in sequence-to-sequence voice conversion models with limited data by proposing pretraining from text-to-speech or automatic speech recognition tasks, resulting in improved intelligibility, naturalness, and similarity, with Transformer-based models outperforming RNN-based ones.
Sequence-to-sequence (seq2seq) voice conversion (VC) models are attractive owing to their ability to convert prosody. Nonetheless, without sufficient data, seq2seq VC models can suffer from unstable training and mispronunciation problems in the converted speech, thus far from practical. To tackle these shortcomings, we propose to transfer knowledge from other speech processing tasks where large-scale corpora are easily available, typically text-to-speech (TTS) and automatic speech recognition (ASR). We argue that VC models initialized with such pretrained ASR or TTS model parameters can generate effective hidden representations for high-fidelity, highly intelligible converted speech. We apply such techniques to recurrent neural network (RNN)-based and Transformer based models, and through systematical experiments, we demonstrate the effectiveness of the pretraining scheme and the superiority of Transformer based models over RNN-based models in terms of intelligibility, naturalness, and similarity.