qwen3-tts-salt-lug-0001

patrickcmd/qwen3-tts-salt-lug-0001

Single-speaker lug finetune of Qwen/Qwen3-TTS-12Hz-1.7B-Base on the salt_lug_0001 voice from Sunbird/tts.

Training summary

• Base model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
• Dataset: Sunbird/tts (config: lug), filtered to speaker_id == salt_lug_0001
• Splits used: train (n=2395), dev (n=50)
• Best dev loss: 6.1009 @ step 575
• Hardware: 1× RTX A6000 48GB
• MLflow run: https://mlflow.sunbird.ai/#/experiments/0/runs/8725586dacb741a0af33d80f24d90b92

Usage

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

tts = Qwen3TTSModel.from_pretrained(
    "patrickcmd/qwen3-tts-salt-lug-0001",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)
wavs, sr = tts.generate_custom_voice(
    text="Oli otya?",
    speaker="salt_lug_0001",
)
sf.write("out.wav", wavs[0], sr)

Limitations

• Single speaker only (salt_lug_0001); voice cloning to other speakers is not the goal of this finetune.
• Trained for 1 epoch on a small subset; expect quality to vary on out-of-distribution lug text.