Tohirju/chatterbox-mtl-tajik

Chatterbox Multilingual — Tajik (t3 fine-tune)

Chatterbox Multilingual v3 fine-tuned for Tajik 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 8,000 (winner of a WER sweep over checkpoints)
Data	100k utterances / 205.5 h Tajik subset of our 642k-utterance Tajik pool
Language token	ru (Tajik is out-of-vocab for Chatterbox; the Russian token fits Tajik Cyrillic best)

Results

Round-trip WER on FLEURS-tg (n=60): synthesize the reference text, transcribe with our Tajik ASR scorer (Whisper-large-v3-tg, checkpoint-14450), score against the reference.

Model	Mean WER	Median WER
This model (Chatterbox-mtl-tajik, step 8000)	10.7%	8.3%
CosyVoice3-tajik-642k	10.8%	7.7%
Orpheus-3B Tajik	~15.9%	—
CosyVoice2 Tajik	16.7%	—

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-tajik")

# 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(
    "Салом! Ин як намунаи овозӣ бо забони тоҷикӣ аст.",
    language_id="ru",                  # Tajik uses the "ru" 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="ru" — the model was fine-tuned under the Russian language token.
• With newer transformers versions the loader needs eager attention (cfg._attn_implementation = "eager").

Files

File	Role
t3_mtl23ls_v2.safetensors	Fine-tuned t3 (Tajik)
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: muhtasham/tajik-audio and the Peacockery Tajik corpus (CC-BY-4.0). Downstream users should observe CC-BY-4.0 attribution for the training data sources.

Citation

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