BlueTTSV2 — Multilingual text-to-speech With VC

This Space: https://huggingface.co/spaces/notmax123/BlueV2

Interactive demo for BlueTTS, running the v2.5 checkpoint from notmax123/BlueTTS2.5-onnx through ONNX Runtime (CUDA where available, CPU otherwise). Project site: lightbluetts.com.

The PyTorch path is still in the tree but is off by default: BLUE_PT_REPO is versioned separately from the ONNX bundle, and voice styles and latent stats are checkpoint-specific, so mixing them degrades quality without erroring. Set BLUE_PT_MATCHES_ONNX_REPO to the PT repo id once a matching v2.5 checkpoint is published.

This Space was duplicated from thewh1teagle/lightblue-tts-hebrew and adapted for the current BlueTTS ONNX bundle.

Languages

• English, Spanish, German, Italian — phonemization with espeak-ng (via phonemizer), aligned with the multilingual tokenizer in the app.
• Hebrew — grapheme-to-phoneme via RenikudPlus (ONNX), conditioned on the selected voice's gender so forms like היא רצה resolve correctly.
• Russian — RUAccent (stress + ё restoration) then RUPhon (stress-conditioned vowel reduction), via the front end bundled at g2p/russian_g2p.py. This is the front end the checkpoint was trained with; espeak-ng's ru voice cannot disambiguate homographs, cannot restore written-out ё, and collides ы with the German ü already in this vocab. Install with pip install -e '.[russian]' — RUAccent/RUPhon pin transformers<5, so they are optional and every other language works without them.

The checkpoint was also trained on Yiddish, which is not exposed here: its front end (nikud + yiddish_g2p) does not ship with the export.

Voices

Four built-in speaker styles (pre-extracted style JSON):

LibriTTS-R references are CC BY 4.0 (Google LLC).

Upload a reference clip to clone a voice. Cloning encodes the clip to latents with the PyTorch autoencoder and then runs the bundle's own reference_encoder.onnx, so the resulting style always matches the acoustic model actually doing the synthesis.

Implementation note

This Space runs inference from app.py (vocabulary, phonemization, ONNX sessions). It does not install the upstream blue Python package from GitHub (that package name replaced the old lightblue-onnx wheel and would conflict with pip’s expected name in a flat requirements.txt).

Model weights

ONNX checkpoints are not stored in this Space repo (file-size limits). On first startup, download_models.py fetches them from the public model repo notmax123/BlueTTS2.5-onnx into ./onnx_slim/ (see ONNX_DIR in app.py), and the Hebrew G2P weights from notmax123/RenikudPlus.

The bundle includes stats.npz and uncond.npz alongside the graphs, and both are load-bearing: v2.5 exports the bare vocoder, so latents are denormalized and folded by the caller, and the vector estimator no longer takes a cfg_scale input, so guidance runs off the unconditional embeddings. A v2 bundle is not a drop-in substitute — download_models.ONNX_BUNDLE_STAMP forces a re-fetch when it changes.

Training / finetuning / voice-export weights (PyTorch / Safetensors): notmax123/blue.

If you fork this Space and point downloads at a private repo, set a Hugging Face write token as the Space secret HF_TOKEN so authenticated downloads still work.

Links

• This demo (Space): huggingface.co/spaces/notmax123/Blue
• ONNX weights: notmax123/BlueTTS2.5-onnx
• PyTorch / training weights: notmax123/blue
• Source code: maxmelichov/BlueTTS on GitHub
• Hebrew G2P: RenikudPlus · notmax123/RenikudPlus

Local development

Dependencies are listed in requirements.txt. With Python 3.12:

pip install -r requirements.txt
python download_models.py   # once, to populate onnx_slim/
python app.py               # or: gradio will launch from app_file when run by HF

To refresh requirements.txt from uv:

uv export --no-hashes --no-emit-project > requirements.txt

License

See the BlueTTS repository and the BlueTTS2.5-onnx model card for licensing of code and weights.