chatterbox-persian-onnx

Chatterbox Persian/Farsi ONNX

Fully controllable Gradio Space for the Persian Chatterbox T3 fine-tune.

The Space contains:

• app.py — Gradio UI and inference controller;
• models/persian_onnx/ — Persian ONNX graphs and external weights;
• src/inference_onnx.py — reusable ONNX Runtime engine;
• src/convert_to_onnx.py — reproducible conversion code;
• src/verify_onnx.py — structural and smoke verification;
• guideline.md — complete conversion and inference guide.

Controls

The UI exposes text, reference voice, INT8/FP32 model choice, exaggeration, CFG, temperature, top-p, min-p, repetition penalty, minimum/maximum speech tokens, seed, CPU thread count, forced CPU, and greedy diagnostic mode. Every result includes a JSON timing report.

The bundled default voice is used when no reference recording is supplied. For best voice cloning, upload a clean mono recording of 3–10 seconds.

License and usage

The Persian checkpoint is reported as CC-BY-NC-4.0. Review the Persian model card and all base-model terms before publishing or using the Space commercially. The model is intended for legitimate, consent-based voice cloning only.

Local upload

From the original repository root:

export HF_TOKEN=hf_...
python chatter_fa/upload_to_hf.py \
  --target space \
  --repo-id YOUR_USER/chatterbox-persian-onnx \
  --include-demo

The upload script excludes caches, reference.wav, and unrelated generated experiments. The model files are large and use Hugging Face storage/LFS.