Modelpublic

Abduqayum/uzbek-tts-natural-speech-chatterbox

sourceHugging Facemitupdated 2mo agoView on Hugging Face

2likes73downloads

Model Card

Uzbek Chatterbox TTS (LoRA fine-tune)

A LoRA fine-tune of ResembleAI/chatterbox (Standard/non-Turbo) for Uzbek speech synthesis, trained on ~30 hours of single-speaker Uzbek audiobook narration.

The LoRA adapter has been merged into the base model, so these are standalone T3 weights — no PEFT required at inference time. The model also supports zero-shot voice cloning: pass any reference wav and it will speak your text in that voice.

Demo

Fixed-voice synthesis (the trained speaker)

Voice cloning (unrelated reference speakers, same fine-tuned model)

Bobur Akmalov <audio controls src="https://huggingface.co/Abduqayum/uzbek-tts-natural-speech-chatterbox/resolve/main/samples/voicecloningdemo/Bobur%20Akmalov.wav"></audio>

Davron Fayziyev <audio controls src="https://huggingface.co/Abduqayum/uzbek-tts-natural-speech-chatterbox/resolve/main/samples/voicecloningdemo/Davron%20Fayziyev.wav"></audio>

Jennifer Lawrence <audio controls src="https://huggingface.co/Abduqayum/uzbek-tts-natural-speech-chatterbox/resolve/main/samples/voicecloningdemo/Jennifer%20Lawrence.wav"></audio>

Files

File	Description
`t3_finetuned_merged.safetensors`	Standalone fine-tuned T3 (speech-token AR model), LoRA merged in
`tokenizer.json`	Grapheme tokenizer, 2454 tokens (unchanged from base — Standard mode uses a fixed multilingual grapheme vocab that already covers Uzbek Latin script)
`reference_voice.wav`	Default reference clip for the fixed training voice
`samples/`	Demo audio (see above)
`inference.py`	Standalone script to run synthesis, fixed-voice or cloning

The s3gen (vocoder) and ve (voice encoder) weights are not re-uploaded here since they're unmodified from the base model — inference.py expects them from a base install (see Usage).

Usage

This model was trained with, and is meant to be run through, gokhaneraslan/chatterbox-finetuning — the same toolkit used to train it. That repo vendors the Chatterbox model code and gives you s3gen.safetensors / ve.safetensors / conds.pt via its setup.py.

bash

git clone https://github.com/gokhaneraslan/chatterbox-finetuning.git
cd chatterbox-finetuning
pip install -r requirements.txt
# if you hit a pkg_resources / setuptools ImportError from resemble-perth:
pip install "setuptools<81"

python setup.py   # downloads base s3gen/ve/conds + the 2454-token tokenizer

# download this repo's files into the toolkit root (or anywhere, and pass paths as flags):
huggingface-cli download <this-repo-id> --local-dir uzbek_chatterbox
cp uzbek_chatterbox/inference.py .

Fixed voice:

bash

python inference.py \
  --text "Salom, bugungi kunda sun'iy intellekt texnologiyalari juda tez rivojlanmoqda." \
  --audio_prompt uzbek_chatterbox/reference_voice.wav \
  --merged_weights uzbek_chatterbox/t3_finetuned_merged.safetensors \
  --output out.wav

Voice cloning (swap in any reference wav):

bash

python inference.py \
  --text "Salom, qandaysiz?" \
  --audio_prompt path/to/some_other_voice.wav \
  --merged_weights uzbek_chatterbox/t3_finetuned_merged.safetensors \
  --output out.wav

Text of any length is handled — inference.py splits long input into sentence-sized chunks internally (a single generate() call is capped at 1000 speech tokens and can otherwise cut off mid-word on long paragraphs).

Training

See TRAINING.md for the full step-by-step recipe (dataset format, exact config, commands) to reproduce this fine-tune on your own dataset/language.

License

MIT, inherited from the base ResembleAI/chatterbox model. The training toolkit (gokhaneraslan/chatterbox-finetuning) is Apache 2.0.