---
license: mit
library_name: pytorch
tags:
  - text-to-speech
  - tts
  - hebrew
  - multilingual
  - flow-matching
  - pytorch
pipeline_tag: text-to-speech
---

# BlueV3 — PyTorch Weights

Multilingual neural TTS (Hebrew-first) based on flow matching in a compressed latent space.

**TTS version:** `v1.7.3` · **Sample rate:** 44.1 kHz · **Checkpoint:** `ckpt_step_767000`

## Contents

| Path | Description |
|------|-------------|
| `checkpoints/text2latent/ckpt_step_767000.pt` | Text encoder + vector-field estimator + reference encoder |
| `checkpoints/duration_predictor/duration_predictor_final.pt` | Utterance duration predictor |
| `stats_multilingual.pt` | Latent mean / std for normalization |
| `configs/tts.json` | Model / AE / DP config |

> **Not included:** the autoencoder / vocoder codec (`ae_*.pt`). Use your local AE checkpoint (e.g. `checkpoints/44k_decoder/ae_541000.pt`) or the ONNX vocoder from [`notmax123/BlueV3-onnx`](https://huggingface.co/notmax123/BlueV3-onnx).

## Companion repos

- ONNX runtime bundle (incl. vocoder): [`notmax123/BlueV3-onnx`](https://huggingface.co/notmax123/BlueV3-onnx)

## Quick start (PyTorch)

```bash
# from the BlueV3 code repo
hf download notmax123/BlueV3 --local-dir ./hf_bluev3

# place weights where run_pt_inference / synth scripts expect them, e.g.:
mkdir -p checkpoints/text2latent checkpoints/duration_predictor
cp hf_bluev3/checkpoints/text2latent/ckpt_step_767000.pt checkpoints/text2latent/
cp hf_bluev3/checkpoints/duration_predictor/duration_predictor_final.pt checkpoints/duration_predictor/
cp hf_bluev3/stats_multilingual.pt .
cp hf_bluev3/configs/tts.json configs/tts.json
```

Example synthesis (IPA / phonemes, Hebrew):

```bash
uv run python run_pt_inference.py \
  --text "metsujˈan. vetatˈus levˈad?" \
  --lang he \
  --style_json voice_styles/Rotem.json \
  --steps 8 \
  --cfg 3.0 \
  --out out.wav
```

You still need:

1. Code from this project’s Git repo  
2. An AE codec checkpoint for waveform decode (or use ONNX vocoder)  
3. A voice style JSON (`style_ttl` + `style_dp`), e.g. export via `export_ref_latent.py` / reference WAV encoding  

## Model notes

- Latent: 24-dim AE → compressed **144** channels (`chunk_compress_factor=6`)
- Flow matching with classifier-free guidance (`cfg_scale` typically `3.0`)
- Universal IPA vocab (size 256)
- Duration is predicted in **seconds**, then mapped to latent frames

## License

MIT (see frontmatter).