🤗 Hugging Face | 🤖 ModelScope | 📑 Blog | 📑 Paper
🖥️ Hugging Face Demo | 🖥️ ModelScope Demo | 💬 WeChat (微信) | 🫨 Discord | 📑 API
We release Qwen3-TTS, a series of powerful speech generation capabilities developed by Qwen, offering comprehensive support for voice clone, voice design, ultra-high-quality human-like speech generation, and natural language-based voice control. It provides developers and users with the most extensive set of speech generation features available.
- 2026.1.22: 🎉🎉🎉 We have released Qwen3-TTS series (0.6B/1.7B) based on Qwen3-TTS-Tokenizer-12Hz. Please check our blog!
Qwen3-TTS covers 10 major languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian) as well as multiple dialectal voice profiles to meet global application needs. In addition, the models feature strong contextual understanding, enabling adaptive control of tone, speaking rate, and emotional expression based on instructions and text semantics, and they show markedly improved robustness to noisy input text. Key features:
- Powerful Speech Representation: Powered by the self-developed Qwen3-TTS-Tokenizer-12Hz, it achieves efficient acoustic compression and high-dimensional semantic modeling of speech signals. It fully preserves paralinguistic information and acoustic environmental features, enabling high-speed, high-fidelity speech reconstruction through a lightweight non-DiT architecture.
- Universal End-to-End Architecture: Utilizing a discrete multi-codebook LM architecture, it realizes full-information end-to-end speech modeling. This completely bypasses the information bottlenecks and cascading errors inherent in traditional LM+DiT schemes, significantly enhancing the model’s versatility, generation efficiency, and performance ceiling.
- Extreme Low-Latency Streaming Generation: Based on the innovative Dual-Track hybrid streaming generation architecture, a single model supports both streaming and non-streaming generation. It can output the first audio packet immediately after a single character is input, with end-to-end synthesis latency as low as 97ms, meeting the rigorous demands of real-time interactive scenarios.
- Intelligent Text Understanding and Voice Control: Supports speech generation driven by natural language instructions, allowing for flexible control over multi-dimensional acoustic attributes such as timbre, emotion, and prosody. By deeply integrating text semantic understanding, the model adaptively adjusts tone, rhythm, and emotional expression, achieving lifelike “what you imagine is what you hear” output.
Below is an introduction and download information for the Qwen3-TTS models that have already been released. Other models mentioned in the technical report will be released in the near future. Please select and download the model that fits your needs.
| Tokenizer Name | Description |
|---|---|
| Qwen3-TTS-Tokenizer-12Hz | The Qwen3-TTS-Tokenizer-12Hz model which can encode the input speech into codes and decode them back into speech. |
| Model | Features | Language Support | Streaming | Instruction Control |
|---|---|---|---|---|
| Qwen3-TTS-12Hz-1.7B-VoiceDesign | Performs voice design based on user-provided descriptions. | Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | ✅ | ✅ |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | Provides style control over target timbres via user instructions; supports 9 premium timbres covering various combinations of gender, age, language, and dialect. | Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | ✅ | ✅ |
| Qwen3-TTS-12Hz-1.7B-Base | Base model capable of 3-second rapid voice clone from user audio input; can be used for fine-tuning (FT) other models. | Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | ✅ | |
| Qwen3-TTS-12Hz-0.6B-CustomVoice | Supports 9 premium timbres covering various combinations of gender, age, language, and dialect. | Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | ✅ | |
| Qwen3-TTS-12Hz-0.6B-Base | Base model capable of 3-second rapid voice clone from user audio input; can be used for fine-tuning (FT) other models. | Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | ✅ |
During model loading in the qwen-tts package or vLLM, model weights will be automatically downloaded based on the model name. However, if your runtime environment is not conducive to downloading weights during execution, you can refer to the following commands to manually download the model weights to a local directory:
# Download through ModelScope (recommended for users in Mainland China)
pip install -U modelscope
modelscope download --model Qwen/Qwen3-TTS-Tokenizer-12Hz --local_dir ./models/Qwen3-TTS-Tokenizer-12Hz
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local_dir ./models/Qwen3-TTS-12Hz-1.7B-CustomVoice
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --local_dir ./models/Qwen3-TTS-12Hz-1.7B-VoiceDesign
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-Base --local_dir ./models/Qwen3-TTS-12Hz-1.7B-Base
modelscope download --model Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice --local_dir ./models/Qwen3-TTS-12Hz-0.6B-CustomVoice
modelscope download --model Qwen/Qwen3-TTS-12Hz-0.6B-Base --local_dir ./models/Qwen3-TTS-12Hz-0.6B-Base
# Download through Hugging Face
pip install -U "huggingface_hub[cli]"
huggingface-cli download Qwen/Qwen3-TTS-Tokenizer-12Hz --local-dir ./models/Qwen3-TTS-Tokenizer-12Hz
huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local-dir ./models/Qwen3-TTS-12Hz-1.7B-CustomVoice
huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --local-dir ./models/Qwen3-TTS-12Hz-1.7B-VoiceDesign
huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-Base --local-dir ./models/Qwen3-TTS-12Hz-1.7B-Base
huggingface-cli download Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice --local-dir ./models/Qwen3-TTS-12Hz-0.6B-CustomVoice
huggingface-cli download Qwen/Qwen3-TTS-12Hz-0.6B-Base --local-dir ./models/Qwen3-TTS-12Hz-0.6B-BaseWhen running from the project root, the Python loaders and demo CLI prefer the local ./models directory. For example, Qwen/Qwen3-TTS-12Hz-1.7B-Base will first resolve to ./models/Qwen3-TTS-12Hz-1.7B-Base if that folder exists; simple names like Qwen3-TTS-12Hz-1.7B-Base are also resolved from ./models by default. The FastAPI service is stricter: API requests still need the supported full official modelId values, but those ids are resolved against the local ./models/<model-leaf> directories.
The easiest way to quickly use Qwen3-TTS is to install the qwen-tts Python package from PyPI after preparing a matching PyTorch environment for your machine. We recommend using a fresh, isolated environment to avoid dependency conflicts with existing packages. You can create a clean Python 3.12 environment like this:
conda create -n qwen3-tts python=3.12 -y
conda activate qwen3-ttsThen install a PyTorch build that matches your machine from the official selector:
After PyTorch is installed, run:
pip install -U qwen-tts accelerateIf you want to develop or modify the code locally, install from source in editable mode.
git clone https://github.com/QwenLM/Qwen3-TTS.git
cd Qwen3-TTS
pip install -e ".[runtime]"If you want to run the API server from source, install the API dependencies as well:
pip install -e ".[runtime,api]"If you also want to use POST /api/translate, install the ASR dependencies:
pip install -e ".[runtime,api,asr]"For the translate API, the project now prefers local ASR model directories under ./models/asr when they exist:
./models/asr/faster-whisper/<modelSize>
Recommended manual download examples:
# faster-whisper medium (ModelScope)
pip install -U modelscope
modelscope download --model Systran/faster-whisper-medium \
--local_dir ./models/asr/faster-whisper/medium
# faster-whisper medium (Hugging Face)
pip install -U "huggingface_hub[cli]"
huggingface-cli download Systran/faster-whisper-medium \
--local-dir ./models/asr/faster-whisper/medium
# faster-whisper large-v2 (ModelScope)
modelscope download --model Systran/faster-whisper-large-v2 \
--local_dir ./models/asr/faster-whisper/large-v2
# faster-whisper large-v2 (Hugging Face)
huggingface-cli download Systran/faster-whisper-large-v2 \
--local-dir ./models/asr/faster-whisper/large-v2
# faster-whisper large-v3 (ModelScope)
modelscope download --model Systran/faster-whisper-large-v3 \
--local_dir ./models/asr/faster-whisper/large-v3
# faster-whisper large-v3 (Hugging Face)
huggingface-cli download Systran/faster-whisper-large-v3 \
--local-dir ./models/asr/faster-whisper/large-v3
# faster-whisper large-v3-turbo (Hugging Face)
huggingface-cli download Systran/faster-whisper-large-v3-turbo \
--local-dir ./models/asr/faster-whisper/large-v3-turboThe included launcher scripts (start_api_linux.sh and start_api_windows.bat) only use the project-local Conda environment at ./py312. If you want to start the API through those scripts, create py312 in the repository root, install a matching PyTorch build first, and then install .[runtime,api]. The scripts ask whether to enable FlashAttention: choose Y to use it, or choose N to start with --no-flash-attn.
Device selection is platform-aware:
- CUDA is used on Linux/Windows/macOS when the current PyTorch build exposes CUDA devices.
- MPS is only used on macOS when
torch.backends.mpsis available. - If neither CUDA nor MPS is available, startup asks whether to continue on CPU.
The API now keeps the control plane in the main process and runs a main-process GPU queue. Each runtime request still gets a fresh worker process, and that worker exits when the request finishes so runtime resources are released between requests. --max-gpu-queue-size defaults to 2, which means up to 2 waiting GPU jobs can be queued behind the active one.
torchaudio is required by the 25Hz tokenizer path. Install a matching PyTorch + torchaudio build from the official PyTorch selector before using this feature.
If API startup keeps the default --flash-attn, the server now fails fast when flash_attn is missing or broken and asks you to install it in the current environment instead of deferring that error until the first generation request.
Additionally, we recommend using FlashAttention 2 to reduce GPU memory usage.
pip install -U flash-attn --no-build-isolationIf your machine has less than 96GB of RAM and lots of CPU cores, run:
MAX_JOBS=4 pip install -U flash-attn --no-build-isolationAlso, you should have hardware that is compatible with FlashAttention 2. Read more about it in the official documentation of the FlashAttention repository. FlashAttention 2 can only be used when a model is loaded in torch.float16 or torch.bfloat16.
After installation, you can import Qwen3TTSModel to run custom voice TTS, voice design, and voice clone. The model weights can be specified either as a Hugging Face model id (recommended) or as a local directory path you downloaded. For all the generate_* functions below, besides the parameters shown and explicitly documented, you can also pass generation kwargs supported by Hugging Face Transformers model.generate, e.g., max_new_tokens, top_p, etc.
For custom voice models (Qwen3-TTS-12Hz-1.7B/0.6B-CustomVoice), you just need to call generate_custom_voice, passing a single string or a batch list, along with language, speaker, and optional instruct. You can also call model.get_supported_speakers() and model.get_supported_languages() to see which speakers and languages the current model supports.
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
# single inference
wavs, sr = model.generate_custom_voice(
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。",
language="Chinese", # Pass `Auto` (or omit) for auto language adaptive; if the target language is known, set it explicitly.
speaker="Vivian",
instruct="用特别愤怒的语气说", # Omit if not needed.
)
sf.write("output_custom_voice.wav", wavs[0], sr)
# batch inference
wavs, sr = model.generate_custom_voice(
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。",
"She said she would be here by noon."
],
language=["Chinese", "English"],
speaker=["Vivian", "Ryan"],
instruct=["", "Very happy."]
)
sf.write("output_custom_voice_1.wav", wavs[0], sr)
sf.write("output_custom_voice_2.wav", wavs[1], sr)For Qwen3-TTS-12Hz-1.7B/0.6B-CustomVoice models, the supported speaker list and speaker descriptions are provided below. We recommend using each speaker’s native language for the best quality. Of course, each speaker can speak any language supported by the model.
| 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 a low, mellow timbre. | Chinese |
| Dylan | Youthful Beijing male voice with a clear, natural timbre. | Chinese (Beijing Dialect) |
| Eric | Lively Chengdu male voice with a slightly husky brightness. | Chinese (Sichuan Dialect) |
| Ryan | Dynamic male voice with strong rhythmic drive. | English |
| Aiden | Sunny American male voice with a clear midrange. | English |
| Ono_Anna | Playful Japanese female voice with a light, nimble timbre. | Japanese |
| Sohee | Warm Korean female voice with rich emotion. | Korean |
For the voice design model (Qwen3-TTS-12Hz-1.7B-VoiceDesign), you can use generate_voice_design to provide the target text and a natural-language instruct description.
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
# single inference
wavs, sr = model.generate_voice_design(
text="哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
language="Chinese",
instruct="体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
)
sf.write("output_voice_design.wav", wavs[0], sr)
# batch inference
wavs, sr = model.generate_voice_design(
text=[
"哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
"It's in the top drawer... wait, it's empty? No way, that's impossible! I'm sure I put it there!"
],
language=["Chinese", "English"],
instruct=[
"体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
"Speak in an incredulous tone, but with a hint of panic beginning to creep into your voice."
]
)
sf.write("output_voice_design_1.wav", wavs[0], sr)
sf.write("output_voice_design_2.wav", wavs[1], sr)For the voice clone model (Qwen3-TTS-12Hz-1.7B/0.6B-Base), to clone a voice and synthesize new content, you just need to provide a reference audio clip (ref_audio) along with its transcript (ref_text). ref_audio can be a local file path, a URL, a base64 string, or a (numpy_array, sample_rate) tuple. If you set x_vector_only_mode=True, only the speaker embedding is used so ref_text is not required, but cloning quality may be reduced.
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-Base",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
ref_audio = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone.wav"
ref_text = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."
wavs, sr = model.generate_voice_clone(
text="I am solving the equation: x = [-b ± √(b²-4ac)] / 2a? Nobody can — it's a disaster (◍•͈⌔•͈◍), very sad!",
language="English",
ref_audio=ref_audio,
ref_text=ref_text,
)
sf.write("output_voice_clone.wav", wavs[0], sr)If you need to reuse the same reference prompt across multiple generations (to avoid recomputing prompt features), build it once with create_voice_clone_prompt and pass it via voice_clone_prompt.
prompt_items = model.create_voice_clone_prompt(
ref_audio=ref_audio,
ref_text=ref_text,
x_vector_only_mode=False,
)
wavs, sr = model.generate_voice_clone(
text=["Sentence A.", "Sentence B."],
language=["English", "English"],
voice_clone_prompt=prompt_items,
)
sf.write("output_voice_clone_1.wav", wavs[0], sr)
sf.write("output_voice_clone_2.wav", wavs[1], sr)For more examples of reusable voice clone prompts, batch cloning, and batch inference, please refer to the example codes. With those examples and the generate_voice_clone function description, you can explore more advanced usage patterns.
If you want a designed voice that you can reuse like a cloned speaker, a practical workflow is: (1) use the VoiceDesign model to synthesize a short reference clip that matches your target persona, (2) feed that clip into create_voice_clone_prompt to build a reusable prompt, and then (3) call generate_voice_clone with voice_clone_prompt to generate new content without re-extracting features every time. This is especially useful when you want a consistent character voice across many lines.
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
# create a reference audio in the target style using the VoiceDesign model
design_model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
ref_text = "H-hey! You dropped your... uh... calculus notebook? I mean, I think it's yours? Maybe?"
ref_instruct = "Male, 17 years old, tenor range, gaining confidence - deeper breath support now, though vowels still tighten when nervous"
ref_wavs, sr = design_model.generate_voice_design(
text=ref_text,
language="English",
instruct=ref_instruct
)
sf.write("voice_design_reference.wav", ref_wavs[0], sr)
# build a reusable clone prompt from the voice design reference
clone_model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-Base",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
voice_clone_prompt = clone_model.create_voice_clone_prompt(
ref_audio=(ref_wavs[0], sr), # or "voice_design_reference.wav"
ref_text=ref_text,
)
sentences = [
"No problem! I actually... kinda finished those already? If you want to compare answers or something...",
"What? No! I mean yes but not like... I just think you're... your titration technique is really precise!",
]
# reuse it for multiple single calls
wavs, sr = clone_model.generate_voice_clone(
text=sentences[0],
language="English",
voice_clone_prompt=voice_clone_prompt,
)
sf.write("clone_single_1.wav", wavs[0], sr)
wavs, sr = clone_model.generate_voice_clone(
text=sentences[1],
language="English",
voice_clone_prompt=voice_clone_prompt,
)
sf.write("clone_single_2.wav", wavs[0], sr)
# or batch generate in one call
wavs, sr = clone_model.generate_voice_clone(
text=sentences,
language=["English", "English"],
voice_clone_prompt=voice_clone_prompt,
)
for i, w in enumerate(wavs):
sf.write(f"clone_batch_{i}.wav", w, sr)If you only want to encode and decode audio for transport or training and so on, Qwen3TTSTokenizer supports encode/decode with paths, URLs, numpy waveforms, and dict/list payloads, for example:
import soundfile as sf
from qwen_tts import Qwen3TTSTokenizer
tokenizer = Qwen3TTSTokenizer.from_pretrained(
"Qwen/Qwen3-TTS-Tokenizer-12Hz",
device_map="cuda:0",
)
enc = tokenizer.encode("https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/tokenizer_demo_1.wav")
wavs, sr = tokenizer.decode(enc)
sf.write("decode_output.wav", wavs[0], sr)For more tokenizer examples (including different input formats and batch usage), please refer to the example codes. With those examples and the description for Qwen3TTSTokenizer, you can explore more advanced usage patterns.
To launch the Qwen3-TTS web ui demo, simply install the qwen-tts package and run qwen-tts-demo. Use the command below for help:
qwen-tts-demo --helpTo launch the demo, you can use the following commands:
# CustomVoice model
qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --ip 0.0.0.0 --port 8000
# VoiceDesign model
qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --ip 0.0.0.0 --port 8000
# Base model
qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-Base --ip 0.0.0.0 --port 8000On Windows, you can also use start_demo_windows.bat. It uses the project-local ./py312 environment, prefers the local ./models/Qwen3-TTS-12Hz-1.7B-Base checkpoint if present, asks whether to enable FlashAttention, and defaults to http://127.0.0.1:7860.
On Linux/macOS and other Unix-like systems, you can also use ./start_demo_linux.sh. It uses the project-local ./py312, local model preference, default http://127.0.0.1:7860, and auto-selects cuda:N, mps, or cpu with PyTorch when --device is not provided. The script asks whether to enable FlashAttention; on non-CUDA machines choose N or pass --no-flash-attn.
If you run the explicit qwen-tts-demo ... --port 8000 commands above, open http://<your-ip>:8000. If you use the launcher scripts, open their configured address instead; by default that is http://127.0.0.1:7860.
To avoid browser microphone permission issues after deploying the server, for Base model deployments, it is recommended/required to run the gradio service over HTTPS (especially when accessed remotely or behind modern browsers/gateways). Use --ssl-certfile and --ssl-keyfile to enable HTTPS. First we need to generate a private key and a self-signed cert (valid for 365 days):
openssl req -x509 -newkey rsa:2048 \
-keyout key.pem -out cert.pem \
-days 365 -nodes \
-subj "/CN=localhost"Then run the demo with HTTPS:
qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-Base \
--ip 0.0.0.0 --port 8000 \
--ssl-certfile cert.pem \
--ssl-keyfile key.pem \
--no-ssl-verifyAnd open https://<your-ip>:8000 to experience it. If your browser shows a warning, it’s expected for self-signed certificates. For production, use a real certificate.
To further explore Qwen3-TTS, we encourage you to try our DashScope API for a faster and more efficient experience. For detailed API information and documentation, please refer to the following:
| API Description | API Documentation (Mainland China) | API Documentation (International) |
|---|---|---|
| Real-time API for Qwen3-TTS of custom voice model. | https://help.aliyun.com/zh/model-studio/qwen-tts-realtime | https://www.alibabacloud.com/help/en/model-studio/qwen-tts-realtime |
| Real-time API for Qwen3-TTS of voice clone model. | https://help.aliyun.com/zh/model-studio/qwen-tts-voice-cloning | https://www.alibabacloud.com/help/en/model-studio/qwen-tts-voice-cloning |
| Real-time API for Qwen3-TTS of voice design model. | https://help.aliyun.com/zh/model-studio/qwen-tts-voice-design | https://www.alibabacloud.com/help/en/model-studio/qwen-tts-voice-design |
vLLM officially provides day-0 support for Qwen3-TTS! Welcome to use vLLM-Omni for Qwen3-TTS deployment and inference. For installation and more details, please check vLLM-Omni official documentation. Now only offline inference is supported. Online serving will be supported later, and vLLM-Omni will continue to offer support and optimization for Qwen3-TTS in areas such as inference speed and streaming capabilities.
You can use vLLM-Omni to inference Qwen3-TTS locally, we provide examples in vLLM-Omni repo which can generate audio output:
# git clone https://github.com/vllm-project/vllm-omni.git
# cd vllm-omni/examples/offline_inference/qwen3_tts
# Run a single sample with CustomVoice task
python end2end.py --query-type CustomVoice
# Batch sample (multiple prompts in one run) with CustomVoice task:
python end2end.py --query-type CustomVoice --use-batch-sample
# Run a single sample with VoiceDesign task
python end2end.py --query-type VoiceDesign
# Batch sample (multiple prompts in one run) with VoiceDesign task:
python end2end.py --query-type VoiceDesign --use-batch-sample
# Run a single sample with Base task in icl mode-tag
python end2end.py --query-type Base --mode-tag iclPlease refer to Qwen3-TTS-Finetuning for detailed instructions on fine-tuning Qwen3-TTS.
During evaluation, we ran inference for all models with dtype=torch.bfloat16 and set max_new_tokens=2048. All other sampling parameters used the defaults from the checkpoint’s generate_config.json. For the Seed-Test and InstructTTS-Eval test sets, we set language="auto", while for all other test sets we explicitly passed the corresponding language. The detailed results are shown below.
Speech Generation Benchmarks
Zero-shot speech generation on the Seed-TTS test set. Performance is measured by Word Error Rate (WER, ↓), where lower is better.
| Datasets | Model | Performance | |
|---|---|---|---|
| Content Consistency | |||
| SEED test-zh | test-en |
Seed-TTS (Anastassiou et al., 2024) | 1.12 | 2.25 |
| MaskGCT (Wang et al., 2024) | 2.27 | 2.62 | |
| E2 TTS (Eskimez et al., 2024) | 1.97 | 2.19 | |
| F5-TTS (Chen et al., 2024) | 1.56 | 1.83 | |
| Spark TTS (Wang et al., 2025) | 1.20 | 1.98 | |
| Llasa-8B (Ye et al., 2025b) | 1.59 | 2.97 | |
| KALL-E (Xia et al., 2024) | 0.96 | 1.94 | |
| FireRedTTS 2 (Xie et al., 2025) | 1.14 | 1.95 | |
| CosyVoice 3 (Du et al., 2025) | 0.71 | 1.45 | |
| MiniMax-Speech (Zhang et al., 2025a) | 0.83 | 1.65 | |
| Qwen3-TTS-25Hz-0.6B-Base | 1.18 | 1.64 | |
| Qwen3-TTS-25Hz-1.7B-Base | 1.10 | 1.49 | |
| Qwen3-TTS-12Hz-0.6B-Base | 0.92 | 1.32 | |
| Qwen3-TTS-12Hz-1.7B-Base | 0.77 | 1.24 | |
Multilingual speech generation on the TTS multilingual test set. Performance is measured by Word Error Rate (WER, ↓) for content consistency and Cosine Similarity (SIM, ↑) for speaker similarity.
| Language | Qwen3-TTS-25Hz | Qwen3-TTS-12Hz | MiniMax | ElevenLabs | ||
|---|---|---|---|---|---|---|
| 0.6B-Base | 1.7B-Base | 0.6B-Base | 1.7B-Base | |||
| Content Consistency | ||||||
| Chinese | 1.108 | 0.777 | 1.145 | 0.928 | 2.252 | 16.026 |
| English | 1.048 | 1.014 | 0.836 | 0.934 | 2.164 | 2.339 |
| German | 1.501 | 0.960 | 1.089 | 1.235 | 1.906 | 0.572 |
| Italian | 1.169 | 1.105 | 1.534 | 0.948 | 1.543 | 1.743 |
| Portuguese | 2.046 | 1.778 | 2.254 | 1.526 | 1.877 | 1.331 |
| Spanish | 2.031 | 1.491 | 1.491 | 1.126 | 1.029 | 1.084 |
| Japanese | 4.189 | 5.121 | 6.404 | 3.823 | 3.519 | 10.646 |
| Korean | 2.852 | 2.631 | 1.741 | 1.755 | 1.747 | 1.865 |
| French | 2.852 | 2.631 | 2.931 | 2.858 | 4.099 | 5.216 |
| Russian | 5.957 | 4.535 | 4.458 | 3.212 | 4.281 | 3.878 |
| Speaker Similarity | ||||||
| Chinese | 0.797 | 0.796 | 0.811 | 0.799 | 0.780 | 0.677 |
| English | 0.811 | 0.815 | 0.829 | 0.775 | 0.756 | 0.613 |
| German | 0.749 | 0.737 | 0.769 | 0.775 | 0.733 | 0.614 |
| Italian | 0.722 | 0.718 | 0.792 | 0.817 | 0.699 | 0.579 |
| Portuguese | 0.790 | 0.783 | 0.794 | 0.817 | 0.805 | 0.711 |
| Spanish | 0.732 | 0.731 | 0.812 | 0.814 | 0.762 | 0.615 |
| Japanese | 0.810 | 0.807 | 0.798 | 0.788 | 0.776 | 0.738 |
| Korean | 0.824 | 0.814 | 0.812 | 0.799 | 0.779 | 0.700 |
| French | 0.698 | 0.703 | 0.700 | 0.714 | 0.628 | 0.535 |
| Russian | 0.734 | 0.744 | 0.781 | 0.792 | 0.761 | 0.676 |
Cross-lingual speech generation on the Cross-Lingual benchmark. Performance is measured by Mixed Error Rate (WER for English, CER for others, ↓).
| Task | Qwen3-TTS-25Hz-1.7B-Base | Qwen3-TTS-12Hz-1.7B-Base | CosyVoice3 | CosyVoice2 |
|---|---|---|---|---|
| en-to-zh | 5.66 | 4.77 | 5.09 | 13.5 |
| ja-to-zh | 3.92 | 3.43 | 3.05 | 48.1 |
| ko-to-zh | 1.14 | 1.08 | 1.06 | 7.70 |
| zh-to-en | 2.91 | 2.77 | 2.98 | 6.47 |
| ja-to-en | 3.95 | 3.04 | 4.20 | 17.1 |
| ko-to-en | 3.48 | 3.09 | 4.19 | 11.2 |
| zh-to-ja | 9.29 | 8.40 | 7.08 | 13.1 |
| en-to-ja | 7.74 | 7.21 | 6.80 | 14.9 |
| ko-to-ja | 4.17 | 3.67 | 3.93 | 5.86 |
| zh-to-ko | 8.12 | 4.82 | 14.4 | 24.8 |
| en-to-ko | 6.83 | 5.14 | 5.87 | 21.9 |
| ja-to-ko | 6.86 | 5.59 | 7.92 | 21.5 |
Controllable speech generation on InstructTTSEval. Performance is measured by Attribute Perception and Synthesis accuracy (APS), Description-Speech Consistency (DSD), and Response Precision (RP).
| Type | Model | InstructTTSEval-ZH | InstructTTSEval-EN | ||||
|---|---|---|---|---|---|---|---|
| APS (↑) | DSD (↑) | RP (↑) | APS (↑) | DSD (↑) | RP (↑) | ||
| Target Speaker |
Gemini-flash | 88.2 | 90.9 | 77.3 | 92.3 | 93.8 | 80.1 |
| Gemini-pro | 89.0 | 90.1 | 75.5 | 87.6 | 86.0 | 67.2 | |
| Qwen3TTS-25Hz-1.7B-CustomVoice | 83.1 | 75.0 | 63.0 | 79.0 | 82.8 | 69.3 | |
| Qwen3TTS-12Hz-1.7B-CustomVoice | 83.0 | 77.8 | 61.2 | 77.3 | 77.1 | 63.7 | |
| GPT-4o-mini-tts | 54.9 | 52.3 | 46.0 | 76.4 | 74.3 | 54.8 | |
| Voice Design |
Qwen3TTS-12Hz-1.7B-VD | 85.2 | 81.1 | 65.1 | 82.9 | 82.4 | 68.4 |
| Mimo-Audio-7B-Instruct (Zhang et al., 2025b) | 75.7 | 74.3 | 61.5 | 80.6 | 77.6 | 59.5 | |
| VoiceSculptor (Hu et al., 2026) | 75.7 | 64.7 | 61.5 | - | - | - | |
| Hume | - | - | - | 83.0 | 75.3 | 54.3 | |
| VoxInstruct (Zhou et al., 2024) | 47.5 | 52.3 | 42.6 | 54.9 | 57.0 | 39.3 | |
| Parler-tts-mini (Lyth & King, 2024) | - | - | - | 63.4 | 48.7 | 28.6 | |
| Parler-tts-large (Lyth & King, 2024) | - | - | - | 60.0 | 45.9 | 31.2 | |
| PromptTTS (Guo et al., 2023) | - | - | - | 64.3 | 47.2 | 31.4 | |
| PromptStyle (Liu et al., 2023) | - | - | - | 57.4 | 46.4 | 30.9 | |
Target-Speaker Multilingual Speech Generation on the TTS multilingual test set. Performance is measured by Word Error Rate (WER, ↓).
| Language | Qwen3-TTS-25Hz | Qwen3-TTS-12Hz | GPT-4o-Audio Preview |
||
|---|---|---|---|---|---|
| 0.6B-CustomVoice | 1.7B-CustomVoice | 0.6B-CustomVoice | 1.7B-CustomVoice | ||
| Chinese | 0.874 | 0.708 | 0.944 | 0.903 | 3.519 |
| English | 1.332 | 0.936 | 1.188 | 0.899 | 2.197 |
| German | 0.990 | 0.634 | 2.722 | 1.057 | 1.161 |
| Italian | 1.861 | 1.271 | 2.545 | 1.362 | 1.194 |
| Portuguese | 1.728 | 1.854 | 3.219 | 2.681 | 1.504 |
| Spanish | 1.309 | 1.284 | 1.154 | 1.330 | 4.000 |
| Japanese | 3.875 | 4.518 | 6.877 | 4.924 | 5.001 |
| Korean | 2.202 | 2.274 | 3.053 | 1.741 | 2.763 |
| French | 3.865 | 3.080 | 3.841 | 3.781 | 3.605 |
| Russian | 6.529 | 4.444 | 5.809 | 4.734 | 5.250 |
Long speech generation results. Performance is measured by Word Error Rate (WER, ↓).
| Datasets | Model | Performance | |
|---|---|---|---|
| Content Consistency | |||
| long-zh | long-en | Higgs-Audio-v2 (chunk) (Boson AI, 2025) | 5.505 | 6.917 |
| VibeVoice (Peng et al., 2025) | 22.619 | 1.780 | |
| VoxCPM (Zhou et al., 2025) | 4.835 | 7.474 | |
| Qwen3-TTS-25Hz-1.7B-CustomVoice | 1.517 | 1.225 | |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | 2.356 | 2.812 | |
Speech Tokenizer Benchmarks
Comparison between different supervised semantic speech tokenizers on ASR Task.
| Model | Codebook Size | FPS | C.V. EN | C.V. CN | Fluers EN | Fluers CN |
|---|---|---|---|---|---|---|
| S3 Tokenizer(VQ) (Du et al., 2024a) | 4096 | 50 | 12.06 | 15.38 | - | - |
| S3 Tokenizer(VQ) (Du et al., 2024a) | 4096 | 25 | 11.56 | 18.26 | 7.65 | 5.03 |
| S3 Tokenizer(FSQ) (Du et al., 2024a) | 6561 | 25 | 10.67 | 7.29 | 6.58 | 4.43 |
| Qwen-TTS-Tokenizer-25Hz (Stage 1) | 32768 | 25 | 7.51 | 10.73 | 3.07 | 4.23 |
| Qwen-TTS-Tokenizer-25Hz (Stage 2) | 32768 | 25 | 10.40 | 14.99 | 4.14 | 4.67 |
Comparison between different semantic-related speech tokenizers.
| Model | NQ | Codebook Size | FPS | PESQ_WB | PESQ_NB | STOI | UTMOS | SIM |
|---|---|---|---|---|---|---|---|---|
| SpeechTokenizer (Zhang et al., 2023a) | 8 | 1024 | 50 | 2.60 | 3.05 | 0.92 | 3.90 | 0.85 |
| X-codec (Ye et al., 2025a) | 2 | 1024 | 50 | 2.68 | 3.27 | 0.86 | 4.11 | 0.84 |
| X-codec 2 (Ye et al., 2025b) | 1 | 65536 | 50 | 2.43 | 3.04 | 0.92 | 4.13 | 0.82 |
| XY-Tokenizer (Gong et al., 2025) | 8 | 1024 | 12.5 | 2.41 | 3.00 | 0.91 | 3.98 | 0.83 |
| Mimi (Défossez et al., 2024) | 16 | 2048 | 12.5 | 2.88 | 3.42 | 0.94 | 3.87 | 0.87 |
| FireredTTS 2 Tokenizer (Xie et al., 2025) | 16 | 2048 | 12.5 | 2.73 | 3.28 | 0.94 | 3.88 | 0.87 |
| Qwen-TTS-Tokenizer-12Hz | 16 | 2048 | 12.5 | 3.21 | 3.68 | 0.96 | 4.16 | 0.95 |
If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝 :)
@article{Qwen3-TTS,
title={Qwen3-TTS Technical Report},
author={Hangrui Hu and Xinfa Zhu and Ting He and Dake Guo and Bin Zhang and Xiong Wang and Zhifang Guo and Ziyue Jiang and Hongkun Hao and Zishan Guo and Xinyu Zhang and Pei Zhang and Baosong Yang and Jin Xu and Jingren Zhou and Junyang Lin},
journal={arXiv preprint arXiv:2601.15621},
year={2026}
}See API.md for the API documentation.