Qwen3-TTS-12Hz-0.6B-Uzbek

Full fine-tune of Qwen/Qwen3-TTS-12Hz-0.6B-Base for Uzbek, with voice cloning from a short reference clip (x-vector speaker embedding). English is kept.

• Data: 20 speakers (12 male, 8 female), about 32.6 h from mrmuminov/uzbek_voice, cleaned with Sidon (48 kHz).
• Training: 6 epochs, lr 3e-6, effective batch 32, fp16 autocast with fp32 master weights.
• Checkpoint: epoch 2 of 6 (lowest validation loss 3.2399).

Usage

import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained("mrmuminov/Qwen3-TTS-12Hz-0.6B-Uzbek", device_map="cuda:0", dtype=torch.float32)
wavs, sample_rate = model.generate_voice_clone(
    text="Salom, bugun ob-havo juda yaxshi.",
    language="Auto",
    ref_audio="reference.wav",
    x_vector_only_mode=True,
    non_streaming_mode=True,
)
sf.write("out.wav", wavs[0], sample_rate)

Use non_streaming_mode=True (the model was trained with the non-streaming text layout) and x_vector_only_mode=True. For English text pass language="English".