- Align datasets
- Implement modules
- Training
- End-To-End Synthesizer
- Add Loss CE RVQ
- Subjective Evaluation
- Objective Evaluation
- Demo Page
LibriTTS test clean- ASR WER
whisper large-v2 - Speaker Embedding https://huggingface.co/docs/transformers/model_doc/wavlm#transformers.WavLMForXVector
| Prompt | WER | Speaker cosine Similarity | UtteranceLevel Pitch Mean MAE | UtteranceLevel Pitch Std MAE | UtteranceLevel Duration Diff |
|---|---|---|---|---|---|
| Ground Truth | 0.86 | - | - | - | - |
| 2 Seconds | |||||
| 4 Seconds | |||||
| 6 Seconds | |||||
| 8 Seconds | |||||
| 4 Seconds(PrefixPrompt) | (avg utter duration） |