NaturalSpeechStage
Enums · AiDotNet.TextToSpeech.CodecBased
NaturalSpeech’s training stages (Tan et al. 2022, Algorithm 1 and App. F).
Fields
| Field | Summary |
|---|---|
NaturalSpeechStage Joint |
The main stage: predicted durations, the memory bank (initialize it first with InitializeMemoryBank), the backward and forward losses. |
NaturalSpeechStage PhonemePretraining |
Mixed-phoneme pre-training of the phoneme encoder (masked phoneme and sup-phoneme prediction on text). |
NaturalSpeechStage Tuning |
The tuning stage (the paper’s last 2k epochs): only the end-to-end loss, only the durator trained. |
NaturalSpeechStage Warmup |
The warm-up stage (the paper’s first 1k epochs): durations from monotonic alignment search train the duration predictor and drive the upsampling, no memory bank, the backward loss only. |
int value__ |