notmax123/blue-onnx-v2

Blue ONNX v2

Slim ONNX graph bundle (opset 17) for Blue / Light Blue TTS inference with ONNX Runtime. Same multilingual pipeline as the main blue-onnx package: Hebrew, English, Spanish, Italian, and German (including inline mixed language with XML-style tags in text).

Live demo: Hugging Face Space notmax123/BlueV2 · Site: lightbluetts.com

Files in this repository

File	Role
text_encoder.onnx	Text encoder; inputs include phone IDs and per-voice style (TTL) tensors
vector_estimator.onnx	Flow / vector field with CFG and unconditional latents baked in
vocoder.onnx	Neural codec decoder; output is waveform (stats baked in-graph)
duration_predictor.onnx	Style-conditioned duration; takes style_dp from the voice JSON
tts.json	Runtime config (dims, model version). Pass your ONNX directory to the library; if tts.json sits next to the .onnx files, it is picked up automatically.

Voice style is not inside this repo. Use per-line JSON files (e.g. female1.json) from the BlueTTS repo / companion Hub releases, and pass their paths to load_voice_style or BlueTTS(..., style_json=...).

For Hebrew G2P you need a separate renikud model.onnx next to your app (see the main Blue README).

PyTorch / safetensors weights and export tools live in the notmax123/blue-v2 model repository.

Download

# Hugging Face CLI (e.g. after: pip install huggingface_hub)
hf download notmax123/blue-onnx-v2 --repo-type model --local-dir ./onnx_slim

Or in Python (uses the same huggingface_hub dependency as blue-onnx):

from huggingface_hub import hf_hub_download
hf_hub_download("notmax123/blue-onnx-v2", "text_encoder.onnx", local_dir="onnx_slim", repo_type="model")
# …repeat for the other files, or use snapshot_download

Usage (Python)

pip install blue-onnx

import soundfile as sf
from blue_onnx import BlueTTS

tts = BlueTTS(
    onnx_dir="onnx_slim",       # this folder after download
    style_json="voices/female1.json",  # from the main repo
    renikud_path="model.onnx",   # optional; Hebrew
)

audio, sr = tts.synthesize("Hello, world.", lang="en")
sf.write("out.wav", audio, sr)

Citation and license

This project redistributes model weights and tooling under the same terms as the BlueTTS project (MIT). If you use this in research, see the GitHub README for BibTeX entries to related papers.