qwen3-tts-yuna-lora

Qwen3-TTS LoRA Adapter - Yuna

LoRA fine-tuned adapter for Qwen/Qwen3-TTS-12Hz-1.7B-Base.

Training Details

• Base Model: Qwen3-TTS-12Hz-1.7B-Base
• Method: LoRA (r=16, alpha=32)
• Data: 1 hour Korean audiobook (Albert Camus - The Stranger)
• Hardware: Apple M4 (MPS acceleration)
• Trainable params: 19.2M / 1.94B (0.99%)

Usage

import torch
from peft import PeftModel
from qwen_tts import Qwen3TTSModel
from safetensors.torch import load_file

# Load base model
model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
    dtype=torch.bfloat16,  # or float32 for MPS
    attn_implementation="flash_attention_2",  # or "eager" for MPS
    device_map="cuda",  # or "mps"
)

# Load LoRA adapter
model.model.talker = PeftModel.from_pretrained(
    model.model.talker,
    "tonymustbegreat/qwen3-tts-yuna-lora",
)

# Load speaker embedding
spk = load_file("tonymustbegreat/qwen3-tts-yuna-lora/speaker_embedding.safetensors")
# Inject into codec embedding at position 3000
model.model.talker.model.model.codec_embedding.weight.data[3000] = spk["speaker_embedding"]

# Generate speech
wavs, sr = model.generate_custom_voice(
    text="안녕하세요, 이것은 파인튜닝된 음성입니다.",
    speaker="yuna",
)

Files

• adapter_config.json - LoRA configuration
• adapter_model.safetensors - LoRA weights (77MB)
• speaker_embedding.safetensors - Speaker embedding vector
• tts_config.json - Custom voice TTS config