Chatterbox-MTL Central-Asian (Tajik · Uzbek · Kazakh · Kyrgyz)

How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.

A single text-to-speech model that speaks four Central-Asian languages — Tajik, Uzbek, Kazakh and Kyrgyz — with the target language selected at inference time via a language_id. It is Chatterbox Multilingual v3 fine-tuned jointly on all four languages. Voice cloning is supported: every generation is conditioned on a reference .wav.

Model summary


Base	Chatterbox Multilingual v3 (ResembleAI)
Fine-tuning	Joint multilingual, t3-only (s3gen and ve frozen)
Precision	fp32
Learning rate	1e-5
Vocabulary	extended to 2458 text tokens (stock is 2454) with per-utterance language tokens `[tg]` `[uz]` `[kk]` `[ky]` at ids 2454–2457
Steps trained	20,000
Winner checkpoint	step 10,000 (best joint quality; later checkpoints regress)
License	MIT
Watermark	Perth (Resemble) watermarked output

Only the fine-tuned extended t3 (t3_mtl23ls_v2.safetensors) is trained; s3gen.pt, ve.pt and conds.pt are the frozen Chatterbox components, and grapheme_mtl_merged_expanded_v1.json is the extended 2458-token tokenizer.

Training data

344,031 utterances total:

Language	Utts	Source
Tajik (tg)	100,000	audiobook + diarized
Uzbek (uz)	100,000	uzbekvoice
Kazakh (kk)	100,000	KSC2
Kyrgyz (ky)	7,339 clips, oversampled 6×	CommonVoice 17 (ky) + FLEURS (ky)

FLEURS test sets were held out for tg / uz / kk.

Results (round-trip TTS eval, n = 60 per language)

Generate from held-out text → transcribe with an ASR scorer → compare to reference.

Language	Metric	This combined model	Specialist (single-lang FT)
Tajik	WER	11.8% median	10.7% median
Uzbek	WER	18.2% median	14.8% median
Kazakh	CER	6.9% median	7.4% median (combined is better)
Kyrgyz	—	not scored (no trusted Kyrgyz ASR scorer yet)	—

Notes:

Kazakh is reported as CER (character error rate): Kazakh is agglutinative and a base-whisper WER scorer inflates the number, so CER is the trustworthy metric.
Uzbek is the weak language in joint training — its Latin-minority script destabilizes relative to the Cyrillic-script Tajik/Kazakh.

Raw evidence is in eval/: per_utt_{tg,uz,kk}_n60.jsonl (raw ref/hyp/wer per utterance at step 10000) and scores.txt.

Inference

CRITICAL: set CB_TEXT_VOCAB=2458 before loading. This model has 2458 text tokens; without this env the loader assumes 2454 and fails.

import os
os.environ["CB_TEXT_VOCAB"] = "2458"  # REQUIRED — extended vocab

from chatterbox.mtl_tts import ChatterboxMultilingualTTS

model = ChatterboxMultilingualTTS.from_local("path/to/model_dir", device="cuda")

# eager-attention workaround
model.t3.tfmr.config._attn_implementation = "eager"

wav = model.generate(
    "Матни шумо дар ин ҷо",       # your text
    language_id="tg",            # one of: tg | uz | kk | ky
    audio_prompt_path="ref.wav", # reference voice (cloning) — required
)

language_id selects the spoken language: "tg", "uz", "kk", or "ky".
A reference .wav (audio_prompt_path) is required — this is a voice-cloning TTS.
Output audio is Perth-watermarked.

Files

t3_mtl23ls_v2.safetensors — fine-tuned extended t3 (trained)
s3gen.pt — frozen S3 generator
ve.pt — frozen voice encoder
conds.pt — default conditioning
grapheme_mtl_merged_expanded_v1.json — extended 2458-token tokenizer ([tg][uz][kk][ky])

License & attributions

Model: MIT (following Chatterbox).
Training data attributions: uzbekvoice (CC0), KSC2 (Kazakh Speech Corpus 2), CommonVoice (CC0), FLEURS (CC-BY).