Changelog

Release history and updates for Qwen3-TTS

Latest Release

January 22, 2026 - Qwen3-TTS Initial Release

We are excited to announce the initial release of Qwen3-TTS, a series of powerful speech generation models developed by the Qwen team.

Released Models

All models are based on the Qwen3-TTS-Tokenizer-12Hz:

Qwen3-TTS-12Hz-1.7B-CustomVoice

1.7B parameter model with 9 premium speakers and instruction control

Qwen3-TTS-12Hz-1.7B-VoiceDesign

1.7B parameter model with voice design from natural language descriptions

Qwen3-TTS-12Hz-1.7B-Base

1.7B parameter base model with 3-second voice cloning capability

Qwen3-TTS-12Hz-0.6B-CustomVoice

0.6B parameter efficient model with 9 premium speakers

Qwen3-TTS-12Hz-0.6B-Base

0.6B parameter efficient base model with voice cloning

Qwen3-TTS-Tokenizer-12Hz

High-fidelity 12Hz speech tokenizer for encoding and decoding

Key Features

Powerful Speech Representation

• Self-developed Qwen3-TTS-Tokenizer-12Hz
• Efficient acoustic compression and high-dimensional semantic modeling
• Preserves paralinguistic information and acoustic environmental features
• High-speed, high-fidelity speech reconstruction

Universal End-to-End Architecture

• Discrete multi-codebook LM architecture
• Full-information end-to-end speech modeling
• Eliminates information bottlenecks of traditional LM+DiT schemes
• Enhanced versatility, generation efficiency, and performance ceiling

Extreme Low-Latency Streaming

• Dual-Track hybrid streaming generation architecture
• Single model supports both streaming and non-streaming
• First audio packet after single character input
• End-to-end synthesis latency as low as 97ms

Intelligent Voice Control

• Natural language instruction-based voice control
• Flexible control over timbre, emotion, and prosody
• Deep text semantic understanding
• Adaptive tone, rhythm, and emotional expression

Comprehensive Language Support

Supports 10 major languages:

• Chinese (Mandarin)
• English
• Japanese
• Korean
• German
• French
• Russian
• Portuguese
• Spanish
• Italian

Plus Chinese dialects: Beijing, Sichuan

Premium Speakers

CustomVoice models include 9 carefully crafted speakers:

Speaker	Voice Description	Native Language
Vivian	Bright, slightly edgy young female voice	Chinese
Serena	Warm, gentle young female voice	Chinese
Uncle_Fu	Seasoned male voice with low, mellow timbre	Chinese
Dylan	Youthful Beijing male voice, clear and natural	Chinese (Beijing)
Eric	Lively Chengdu male voice, husky brightness	Chinese (Sichuan)
Ryan	Dynamic male voice with strong rhythmic drive	English
Aiden	Sunny American male voice, clear midrange	English
Ono_Anna	Playful Japanese female voice, light and nimble	Japanese
Sohee	Warm Korean female voice with rich emotion	Korean

Performance Highlights

• Best-in-class English synthesis: WER of 1.24 on Seed-TTS test-en
• Competitive Chinese synthesis: WER of 0.77 on Seed-TTS test-zh
• State-of-the-art voice design: Leading performance on InstructTTSEval
• Strong cross-lingual: Excellent results on language transfer tasks
• Multilingual excellence: Competitive performance across all 10 languages

See the Benchmarks page for detailed results.

Resources

Technical Paper

Read the research paper on arXiv

Blog Post

Detailed introduction and use cases

Hugging Face

Download models from Hugging Face

ModelScope

Download models from ModelScope (CN)

GitHub Repository

Source code and examples

Live Demo (HF)

Try Qwen3-TTS in your browser

Installation

# Install via PyPI
pip install qwen-tts

# Or install from source
git clone https://github.com/QwenLM/Qwen3-TTS.git
cd Qwen3-TTS
pip install -e .

# Optional: Install FlashAttention for better performance
pip install flash-attn --no-build-isolation

Quick Start

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

# Load CustomVoice model
model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

# Generate speech
wavs, sr = model.generate_custom_voice(
    text="Hello! Welcome to Qwen3-TTS.",
    language="English",
    speaker="Ryan",
    instruct="Say it with enthusiasm",
)

sf.write("output.wav", wavs[0], sr)

What’s Coming

Additional models mentioned in the technical report will be released in the near future. Stay tuned!

---

Version History

qwen-tts Package

v0.1.1 - Current Release

Package Information:

• Python support: 3.9, 3.10, 3.11, 3.12, 3.13
• License: Apache-2.0
• Command-line tool: qwen-tts-demo

Dependencies:

• transformers 4.57.3
• accelerate 1.12.0
• gradio (latest)
• librosa, torchaudio, soundfile
• onnxruntime, einops

---

Related Releases

DashScope API

Production-ready API access for Qwen3-TTS models:

CustomVoice API

Real-time streaming API

Voice Clone API

Voice cloning API

Voice Design API

Voice design API

vLLM-Omni Support

vLLM officially provides day-0 support for Qwen3-TTS:

• Optimized inference engine
• Better throughput and latency
• Offline inference (online serving coming soon)
• See vLLM-Omni documentation

---

Stay Updated

GitHub Releases

Watch for new releases

Discord Community

Join discussions

WeChat Group

Chinese community

Subscribe to the GitHub repository to get notified about new releases and updates.