Fine-tuning Guide

Learn how to fine-tune Qwen3-TTS models for custom single-speaker applications

Qwen3-TTS-12Hz-1.7B/0.6B-Base models support single-speaker fine-tuning to adapt the model to your specific voice and use case. This guide walks you through the complete fine-tuning workflow.

Multi-speaker fine-tuning and other advanced fine-tuning features will be supported in future releases.

When to Fine-tune

Fine-tuning is recommended when you need:

• Consistent custom voice: Train the model to generate a specific voice character consistently
• Domain adaptation: Improve performance on specialized vocabulary or speaking styles
• Voice quality: Enhance the naturalness and consistency for a particular speaker
• Single-speaker applications: Apps or services that use one primary voice throughout

Prerequisites

Before starting, ensure you have:

1. Installed the qwen-tts package:

pip install qwen-tts

2. Cloned the repository:

git clone https://github.com/QwenLM/Qwen3-TTS.git
cd Qwen3-TTS/finetuning

Dataset Preparation

Input Format

Prepare your training data as a JSONL file (one JSON object per line). Each line must contain three fields:

• audio: Path to the target training audio (WAV format)
• text: Transcript corresponding to the audio
• ref_audio: Path to the reference speaker audio (WAV format)

Example JSONL

{"audio":"./data/utt0001.wav","text":"其实我真的有发现，我是一个特别善于观察别人情绪的人。","ref_audio":"./data/ref.wav"}
{"audio":"./data/utt0002.wav","text":"She said she would be here by noon.","ref_audio":"./data/ref.wav"}

Reference Audio Recommendations

Best practice: Use the same ref_audio for all samples in your dataset.

Keeping ref_audio identical across the dataset usually improves:

• Speaker consistency during generation
• Training stability
• Voice quality in the fine-tuned model

The reference audio should:

• Be 3-10 seconds long
• Contain clear speech from the target speaker
• Have minimal background noise
• Be in WAV format at 24kHz sampling rate

Training Process

Step 1: Extract Audio Codes

Convert your raw JSONL into a training-ready format that includes audio codes:

python prepare_data.py \
  --device cuda:0 \
  --tokenizer_model_path Qwen/Qwen3-TTS-Tokenizer-12Hz \
  --input_jsonl train_raw.jsonl \
  --output_jsonl train_with_codes.jsonl

Parameters:

• --device: GPU device to use (e.g., cuda:0)
• --tokenizer_model_path: Path or model ID of the tokenizer
• --input_jsonl: Path to your raw JSONL file
• --output_jsonl: Output path for processed JSONL with audio codes

This script uses the Qwen3-TTS-Tokenizer to encode audio into discrete codes that the model can train on.

Step 2: Run Fine-tuning

Launch the supervised fine-tuning (SFT) process:

python sft_12hz.py \
  --init_model_path Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --output_model_path output \
  --train_jsonl train_with_codes.jsonl \
  --batch_size 32 \
  --lr 2e-6 \
  --num_epochs 10 \
  --speaker_name speaker_test

Training Parameters:

Parameter	Description	Recommended Value
--init_model_path	Base model to fine-tune	Qwen/Qwen3-TTS-12Hz-1.7B-Base or Qwen/Qwen3-TTS-12Hz-0.6B-Base
--output_model_path	Directory to save checkpoints	output
--train_jsonl	Training data with codes	Path from Step 1
--batch_size	Batch size per GPU	2-32 (depends on GPU memory)
--lr	Learning rate	2e-6 to 2e-5
--num_epochs	Number of training epochs	3-10
--speaker_name	Name for the custom speaker	Any string (e.g., speaker_test)

Checkpoints are saved after each epoch:

output/
├── checkpoint-epoch-0/
├── checkpoint-epoch-1/
├── checkpoint-epoch-2/
└── ...

Step 3: Test Your Model

Quickly test the fine-tuned model with inference:

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

device = "cuda:0"
tts = Qwen3TTSModel.from_pretrained(
    "output/checkpoint-epoch-2",
    device_map=device,
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

wavs, sr = tts.generate_custom_voice(
    text="She said she would be here by noon.",
    speaker="speaker_test",  # Use the same speaker_name from training
)
sf.write("output.wav", wavs[0], sr)

Configuration Options

Hardware Requirements

• GPU Memory: 16GB+ recommended for 1.7B model with batch_size=2
• RAM: 32GB+ recommended
• GPU: NVIDIA GPU with CUDA support, FlashAttention 2 compatible

Training Configuration

The sft_12hz.py script uses the following configuration:

• Mixed precision: bfloat16 for memory efficiency
• Gradient accumulation: 4 steps
• Optimizer: AdamW with weight decay 0.01
• Gradient clipping: Max norm 1.0
• Loss function: Combined main codec loss + 0.3 × sub-talker loss

Data Configuration

In dataset.py, the training pipeline:

• Loads audio at 24kHz sampling rate
• Extracts mel-spectrograms (128 mels, 1024 FFT size)
• Tokenizes text with Qwen3-TTS processor
• Creates dual-channel inputs (text + codec channels)
• Applies speaker embeddings from reference audio

Best Practices

Data Collection

1. Quality over quantity: 30-100 high-quality samples often work better than 1000+ low-quality samples
2. Diverse content: Include varied sentences covering different phonemes and prosody patterns
3. Clean audio: Use professional recordings or well-processed audio with minimal noise
4. Consistent environment: Record all samples in similar acoustic conditions

Training Tips

1. Start with lower learning rate: 2e-6 is safer; increase to 2e-5 if underfitting
2. Monitor loss: Training loss should steadily decrease; if it plateaus early, try:
   • Increasing learning rate
   • Adding more diverse training data
   • Training for more epochs
3. Batch size: Adjust based on GPU memory:
   • 24GB GPU: batch_size 32 for 0.6B, 16 for 1.7B
   • 16GB GPU: batch_size 16 for 0.6B, 2-4 for 1.7B
4. Checkpoint selection: Test multiple epoch checkpoints; later isn’t always better

Evaluation

• Listen to outputs: Subjective quality is most important
• Test diverse inputs: Try short/long sentences, different emotions, edge cases
• Compare to base model: Ensure fine-tuning improved quality for your use case
• Check consistency: Generate same sentence multiple times to verify consistency

One-Click Training Script

For convenience, combine all steps into a single shell script:

#!/usr/bin/env bash
set -e

DEVICE="cuda:0"
TOKENIZER_MODEL_PATH="Qwen/Qwen3-TTS-Tokenizer-12Hz"
INIT_MODEL_PATH="Qwen/Qwen3-TTS-12Hz-1.7B-Base"

RAW_JSONL="train_raw.jsonl"
TRAIN_JSONL="train_with_codes.jsonl"
OUTPUT_DIR="output"

BATCH_SIZE=2
LR=2e-5
EPOCHS=3
SPEAKER_NAME="speaker_1"

# Step 1: Prepare data
python prepare_data.py \
  --device ${DEVICE} \
  --tokenizer_model_path ${TOKENIZER_MODEL_PATH} \
  --input_jsonl ${RAW_JSONL} \
  --output_jsonl ${TRAIN_JSONL}

# Step 2: Fine-tune
python sft_12hz.py \
  --init_model_path ${INIT_MODEL_PATH} \
  --output_model_path ${OUTPUT_DIR} \
  --train_jsonl ${TRAIN_JSONL} \
  --batch_size ${BATCH_SIZE} \
  --lr ${LR} \
  --num_epochs ${EPOCHS} \
  --speaker_name ${SPEAKER_NAME}

Save this as train.sh, make it executable (chmod +x train.sh), and run with ./train.sh.

Troubleshooting

Out of Memory (OOM)

• Reduce --batch_size
• Use the 0.6B model instead of 1.7B
• Enable gradient checkpointing (modify sft_12hz.py)

Poor Quality Output

• Ensure reference audio is high quality
• Use the same ref_audio for all training samples
• Train for more epochs (try 10-20)
• Check that audio files are 24kHz WAV format

Training Too Slow

• Increase batch size if GPU memory allows
• Use multiple GPUs (requires modifying training script)
• Ensure FlashAttention 2 is installed

Next Steps

• Learn about vLLM integration for optimized inference
• Explore performance optimization techniques
• Try the DashScope API for cloud deployment