ayousanz

/

piper-plus-base

like 8

We are planning to develop a high-quality model that supports both English and Japanese. Due to limited GPU resources required for pre-training, it will take some time before we can release the next model.

This is great news. Thank you. I'm looking forward to it!

@celophi
The quality of Japanese speech has been further improved since the previous version. We have added a perceptual quality discriminator based on Microsoft WavLM to enhance audio quality. Additionally, this update includes improvements to ONNX export for WavLM models (specifically, stochastic sampling and the application of EMA weights).
We also expect to release a mixed Japanese-English model later this year.

@celophi

We are excited to announce that the multilingual base model is now available! This model has been significantly upgraded from the previous Japanese-only version.

What's new:

• 6 languages supported: Japanese, English, Chinese, Spanish, French, and Portuguese
• 571 speakers, 508K utterances used for pre-training (75 epochs, ~282K gradient steps)
• Prosody features (A1/A2/A3) for natural Japanese intonation
• In-sentence code-switching: You can mix languages within a single sentence (e.g., "今日はgood morningですね")
• Fine-tuning friendly: Use --resume-from-multispeaker-checkpoint for single-speaker fine-tuning with automatic Duration Predictor freezing

The model checkpoint (model.ckpt, 302 MB) and updated config/README are already uploaded to this repository. Please see the README for detailed fine-tuning instructions.

We also plan to upload the fine-tuned Tsukuyomi-chan 6-language model to ayousanz/piper-plus-tsukuyomi-chan soon.

Thanks for the update! When I find some time, I will try fine tuning. The in-sentence code code switching is really difficult to achieve for TTS, so nice work!