BlueV3 β PyTorch Weights
Multilingual zero-shot TTS at 44.1 kHz, trained on 7 languages (Hebrew, English, German, Spanish, Italian, Russian, Yiddish): flow-matching text-to-latent model with classifier-free guidance, a duration predictor and an audio autoencoder. Voice identity comes from a reference clip.
Version: v3.0.1 (checkpoint 702000) = text-to-latent EMA at 700k + 1,500 iterations of FlowTTS-GRPO (arXiv 2606.23190) + 500 iterations with a pause reward. Previous release: git tag v3.0.0 (checkpoint 701500).
Operating point: 8 Euler steps, CFG 4.0, speed 1.0, language tags on.
What's new in v3.0.1
- More natural pauses: 500 more RL iterations (FlowTTS-GRPO) with a reward that moves the share of pauses toward read speech; quality and similarity equal or better (table below).
- New duration model: Hebrew is paced slower and closer to read speech.
- Calibrated speed:
speed=1.0now means natural read-speech pace for every voice and language. Each voice JSON carriesspeed_scale(per language); BlueTTS multiplies the requested speed by it. Needs BlueTTSmainfrom 2026-11-15 on; older code ignores the field and plays at the uncalibrated pace. - The previous release (v3.0.0, checkpoint
701500) stays available under the git tagv3.0.0.
Contents
| Path | Description |
|---|---|
checkpoints/text2latent/ckpt_step_702000.pt |
text encoder + vector field + reference encoder (EMA weights) and the unconditional tokens |
checkpoints/duration_predictor/duration_predictor_final.pt |
duration predictor (EMA) |
stats_multilingual.pt |
latent mean / std for this autoencoder (valid only with it) |
configs/tts.json |
model / AE / DP config |
voices/ |
15 voice JSONs exported from this checkpoint: maya, noa, lily, adam, daniel, tamar, shira, yael, emma, sara, david, jonathan, ethan, michael, ariel (LibriTTS-R speakers, CC BY 4.0) |
Autoencoder (codec): notmax123/blue-codec, one file model.safetensors
(EMA 380k). Checkpoints, stats and voices from earlier releases are not compatible with it.
ONNX runtime bundle: notmax123/BlueV3-onnx Β· code:
BlueTTS.
Download
hf download notmax123/BlueV3 --local-dir ./hf_bluev3
hf download notmax123/blue-codec model.safetensors --local-dir ./hf_codec
Scores
Unseen test set (200 Hebrew + 100 English sentences, unseen voices), 8 steps, CFG 4. better = paired 95 % interval excludes zero.
| before RL (700k) | this release | |
|---|---|---|
| Hebrew phone error rate | 1.85 % | 1.53 % |
| Hebrew speaker similarity | 0.650 | 0.660 |
| Hebrew DNSMOS OVRL | 3.39 | 3.41 |
| English word error rate | 1.82 % | 1.88 % (no change) |
| English speaker similarity | 0.660 | 0.671 |
| English DNSMOS OVRL | 3.34 | 3.35 (no change) |
Per-language check
13 voices x 4 sentences per language, 8 steps, CFG 4. Speaker similarity = ECAPA cosine to the voice's own reference (voices speaking a language other than their own score lower); UTMOS = predicted naturalness.
| language | speaker similarity | UTMOS |
|---|---|---|
| Hebrew | 0.496 | 3.54 |
| English | 0.734 | 4.42 |
| German | 0.441 | 3.91 |
| Spanish | 0.454 | 3.76 |
| Italian | 0.432 | 3.78 |
| Russian | 0.402 | 3.74 |
| Yiddish | 0.278 | 3.85 |
Text front ends (G2P) for all 7 languages ship in notmax123/BlueV3-onnx
(frontend/).
The previous release (v1.7.3, ckpt_step_767000, a different architecture) stays available under the git tag v1.7.3.
License
MIT. The codec's decoder is an OpenRAIL-M derivative (see notmax123/blue-codec).