FloWaveNet : A Generative Flow for Raw Audio
share
Most of modern text-to-speech architectures use a WaveNet vocoder for synthesizing a high-fidelity waveform audio, but there has been a limitation for practical applications due to its slow autoregressive sampling scheme. A recently suggested Parallel WaveNet has achieved a real-time audio synthesis by incorporating Inverse Autogressive Flow (IAF) for parallel sampling. However, the Parallel WaveNet requires a two-stage training pipeline with a well-trained teacher network and is prone to mode collapsing if using a probability distillation training only. We propose FloWaveNet, a flow-based generative model for raw audio synthesis. FloWaveNet requires only a single maximum likelihood loss without any additional auxiliary terms and is inherently parallel due to the flow-based transformation. The model can efficiently sample the raw audio in real-time with a clarity comparable to the original WaveNet and ClariNet. Codes and samples for all models including our FloWaveNet is available via GitHub: https://github.com/ksw0306/FloWaveNet
READ FULL TEXT