Reza2kn/Gooya-Bozorg-v1.5

Gooya Bozorg v1.5

Gooya Bozorg v1.5 is the versioned PersianASR distribution of Thomcles/Chatterbox-TTS-Persian-Farsi, a Persian fine-tune of Resemble AI's 500M multilingual Chatterbox TTS model.

This release preserves the selected upstream checkpoint exactly. We did not claim additional training or altered weights: it was promoted after a controlled listening comparison against our own 227,753-clip step-3500 fine-tune, where the upstream Persian checkpoint retained slightly better pronunciation and pacing. The weight hashes are recorded in MODEL_MANIFEST.json.

Why “Bozorg”?

This is the larger, higher-quality Gooya lane. It prioritizes Persian speech quality and zero-shot voice cloning over the footprint of Gooya's small/on-device models.

Quick start

Install the official Chatterbox repository and the small runtime dependencies:

git clone https://github.com/resemble-ai/chatterbox.git
cd chatterbox
pip install -e .
pip install huggingface_hub safetensors soundfile

Then download this repository and run:

python inference.py \
  --model-dir /path/to/Gooya-Bozorg-v1.5 \
  --text "سلام، حالت چطوره؟" \
  --reference /path/to/reference.wav \
  --output gooya-bozorg.wav

The reference should be a clean single-speaker WAV. The defaults match the comparison run: exaggeration=0.5 and cfg_weight=0.5.

Included comparison samples

• samples/canonical.wav: conversational Persian pronunciation
• samples/conversational.wav: punctuation and dialogue-like pacing
• samples/codeswitch.wav: mixed Persian and English terms

All three were generated with fixed seeds and the same reference/settings used for the competing Gooya checkpoint.

Provenance

• Upstream model: Thomcles/Chatterbox-TTS-Persian-Farsi
• Pinned upstream snapshot: 4e9f6b7043d30d8328bb842b6126676bead8d9de
• Architecture/runtime: ResembleAI/chatterbox
• Weight status: byte-identical repackaging of the selected upstream files
• PersianASR release name: Gooya Bozorg v1.5

Credit for the Persian training and original release belongs to Thomcles. Credit for Chatterbox belongs to Resemble AI.

Limitations

• This model can still mispronounce ambiguous Persian words and mixed-script text.
• Pacing and question intonation are improved relative to our tested fine-tune, but are not fully controllable.
• Output quality depends strongly on the reference clip.
• This release is not a tiny/on-device model and needs substantially more memory than Gooya's small models.
• The license is non-commercial; review it before deployment.

License

The upstream checkpoint is released under CC BY-NC 4.0. This repository keeps the same license and attribution requirements. It is not licensed for commercial use.