AdaSpeechTrainingPhase

Enums · AiDotNet.TextToSpeech.Classic

The training phases of AdaSpeech (Chen et al. 2021, §2.3, §3), run in order.

Fields

Field Summary
AdaSpeechTrainingPhase Adaptation Per-voice adaptation: only the speaker embedding and the conditional layer normalizations’ W_γ and W_β (2000 steps).
AdaSpeechTrainingPhase SourceJoint Source-model training together with the phoneme-level acoustic predictor, whose MSE to the stop-gradient phoneme-level encoder output joins the loss (40k steps).
AdaSpeechTrainingPhase SourcePretraining Source-model training of every parameter except the phoneme-level acoustic predictor (60k steps).
int value__