Vocoder-Projected Feature Discriminator
This addresses efficiency bottlenecks in voice conversion and text-to-speech systems for researchers and practitioners, though it is an incremental improvement on existing adversarial training methods.
The paper tackles the computational overhead of waveform-based adversarial training in voice conversion by proposing a vocoder-projected feature discriminator (VPFD) that uses vocoder features instead. This approach achieved comparable performance to waveform discriminators while reducing training time by 9.6× and memory consumption by 11.4× in experiments on diffusion-based voice conversion distillation.
In text-to-speech (TTS) and voice conversion (VC), acoustic features, such as mel spectrograms, are typically used as synthesis or conversion targets owing to their compactness and ease of learning. However, because the ultimate goal is to generate high-quality waveforms, employing a vocoder to convert these features into waveforms and applying adversarial training in the time domain is reasonable. Nevertheless, upsampling the waveform introduces significant time and memory overheads. To address this issue, we propose a vocoder-projected feature discriminator (VPFD), which uses vocoder features for adversarial training. Experiments on diffusion-based VC distillation demonstrated that a pretrained and frozen vocoder feature extractor with a single upsampling step is necessary and sufficient to achieve a VC performance comparable to that of waveform discriminators while reducing the training time and memory consumption by 9.6 and 11.4 times, respectively.