Tohirju/chatterbox-mtl-uzbek

Chatterbox Multilingual — Uzbek (t3 fine-tune)

Chatterbox Multilingual v3 fine-tuned for Uzbek (Latin script) by Saidzoda Lab.

Only the t3 text-to-speech-token transformer was fine-tuned; s3gen and ve (voice encoder) are the frozen upstream weights. The model does voice cloning: a reference wav is required at inference time.

Training

Setting	Value
Base	Chatterbox Multilingual v3 (t3_mtl23ls_v2, stock 2454-token vocab)
Trained modules	t3 only (s3gen / ve frozen)
Precision	fp32
Learning rate	1e-5
Steps trained	12,500
Shipped checkpoint	step 6,000 (winner of a WER sweep over checkpoints)
Data	100k utterances / 119.4 h from the uzbekvoice corpus (FLEURS-uz held out — excluded from training)
Language token	tr (Uzbek Latin is out-of-vocab for Chatterbox; the Turkish token fits Uzbek Latin best)

Results

Round-trip WER on FLEURS-uz (n=60): synthesize the reference text, transcribe with our Uzbek ASR scorer (Whisper-CA, Uzbek), score against the reference.

Model	Mean WER	Median WER
This model (Chatterbox-mtl-uzbek, step 6000)	19.1%	14.8%
CosyVoice3-Uzbek (full 645k)	—	16.0%

This Chatterbox fine-tune beats CosyVoice3-Uzbek on median WER (14.8% vs 16.0%).

Per-utterance evidence for this model is in eval/per_utt.jsonl and eval/scores.json.

Usage

Download the repo files into a directory, then load with the chatterbox-tts package (pip install chatterbox-tts, or add a source checkout to sys.path):

import torchaudio
from huggingface_hub import snapshot_download
from chatterbox.mtl_tts import ChatterboxMultilingualTTS

ckpt_dir = snapshot_download("Tohirju/chatterbox-mtl-uzbek")

# Workaround for newer transformers: the loader expects eager attention.
# If loading fails with an attention-implementation error, patch the t3 config
# before/while loading: cfg._attn_implementation = "eager"

model = ChatterboxMultilingualTTS.from_local(ckpt_dir, device="cuda")

wav = model.generate(
    "Assalomu alaykum! Bu oʻzbek tilidagi ovozli namuna.",
    language_id="tr",                  # Uzbek (Latin) uses the "tr" language token
    audio_prompt_path="ref.wav",       # REQUIRED: reference speaker wav (voice cloning)
)
torchaudio.save("out.wav", wav, model.sr)

Notes:

• audio_prompt_path (a reference wav of the target speaker) is required — this is a voice-cloning model.
• Pass language_id="tr" — the model was fine-tuned under the Turkish language token, which is the closest fit for Uzbek Latin.
• With newer transformers versions the loader needs eager attention (cfg._attn_implementation = "eager").
• Generated audio is Perth-watermarked (upstream Chatterbox behaviour).

Files

File	Role
t3_mtl23ls_v2.safetensors	Fine-tuned t3 (Uzbek)
s3gen.pt	Upstream speech-token-to-wav generator (frozen)
ve.pt	Upstream voice encoder (frozen)
conds.pt	Default conditioning
grapheme_mtl_merged_expanded_v1.json	Grapheme tokenizer

License and data

• Model / code: MIT (upstream Chatterbox is MIT).
• Training data: the uzbekvoice community Uzbek read-speech corpus (CC0-1.0). Downstream users should observe the corpus's own license terms.

Citation

Saidzoda Lab, Central-Asian speech program (Tajik / Uzbek / Kazakh / Kyrgyz), 2026.