nur-dev/ait-syn-0.6b-trilingual-v30

AIT-Syn 0.6B Trilingual — v33 Natural

Public, manually gated checkpoint for Kazakh, Russian, and English text-to-speech with voice cloning. Kazakh is the primary language. The Hugging Face repository name retains its historical v30 suffix, but the weights in the current main revision are the selected v33 step-4000 release.

This lineage starts from Qwen/Qwen3-TTS-12Hz-0.6B-Base and continues from the project's selected v30 checkpoint. It does not use weights from the 1.7B model.

Supported languages

Use the full language names expected by Qwen3-TTS:

Language	ISO	Inference value
Kazakh	kk	kazakh
Russian	ru	russian
English	en	english

Kazakh uses codec_language_id.kazakh = 2072 in the model configuration.

Release details

Property	Value
Release	v33 Natural, step 4000
Parameters	914,643,008 total; 764,218,368 trained
Output	24 kHz mono
Primary use	ICL voice cloning using matching reference audio and text
Training rows	603,105
Validation rows	768 seen-speaker utterances
Evaluation rows	600 unseen-speaker utterances
Training hours	approximately 1,140.25
Natural row mix	56.37% Kazakh, 16.29% Russian, 27.34% English
Sampling policy	shuffled without replacement; no language oversampling

The corrected nur-dev/pronto-speach-fixed source was used once in the Kazakh corpus. The aliased Farabi and older Pronto copies were excluded rather than duplicated.

The selected checkpoint's validation codec losses were 1.2403 for Kazakh, 1.1978 for Russian, and 0.7328 for English (macro 1.05695). A deterministic three-language generation smoke test produced valid, bounded audio in all three languages. In controlled Kazakh listening comparisons, step 4000 was preferred over the previous v30 checkpoint and earlier v33 checkpoints.

Installation

pip install torch soundfile qwen-tts huggingface_hub
# Optional on supported GPUs:
pip install flash-attn --no-build-isolation

Authenticate before loading this manually gated repository:

huggingface-cli login

Voice cloning

ICL voice cloning is the recommended path. ref_text must be an exact transcript of ref_audio.

import soundfile as sf
import torch
from qwen_tts.inference.qwen3_tts_model import Qwen3TTSModel

MODEL_ID = "nur-dev/ait-syn-0.6b-trilingual-v30"

try:
    import flash_attn  # noqa: F401
    attention = "flash_attention_2"
except Exception:
    attention = "eager"

model = Qwen3TTSModel.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    attn_implementation=attention,
    device_map="cuda:0",
)
model.model.eval()

wavs, sample_rate = model.generate_voice_clone(
    text="Сәлеметсіз бе, бұл қазақ тіліндегі дауыс клондау мысалы.",
    ref_audio="reference.wav",
    ref_text="Референс жазбаның дәл мәтіні.",
    language="kazakh",
    x_vector_only_mode=False,
    non_streaming_mode=True,
    do_sample=True,
    temperature=0.7,
    top_k=30,
    top_p=0.95,
    max_new_tokens=8192,
)
sf.write("kazakh_clone.wav", wavs[0], sample_rate, subtype="PCM_16")

Use language="russian" or language="english" for the other languages. Transcript-free cloning is available with x_vector_only_mode=True, but ICL mode usually retains the voice more reliably.

Recommended decoding

For Kazakh production synthesis, start with:

do_sample=True
temperature=0.7
top_k=30
top_p=0.95

Greedy decoding (do_sample=False) is useful for deterministic regression tests, but listening tests found it slightly more mechanical in Kazakh. The 0.8/50/0.95 preset can add variation but is less conservative.

Practical guidance

• Use clean, single-speaker reference audio of roughly 3–10 seconds.
• Supply the exact reference transcript for ICL cloning.
• Use native-language reference material when possible.
• Pin a Hugging Face commit for production deployments because main is the rolling release branch.
• Install FlashAttention for practical inference latency on compatible GPUs.

Limitations and safety

The release passed a small trilingual generation/listening smoke test, not a full human MOS study. Kazakh prosody is improved over the previous release but can still sound slightly less natural than Russian or English on some prompts. Evaluate multiple speakers and sentence styles before deployment.

Do not use the model to impersonate a person without consent. Collect and use reference voices with permission.