whatmodel.sh / tts / Qwen3-TTS 12Hz 1.7B Base
Qwen3-TTS 12Hz 1.7B Base
↗ share
+ contribute
Qwen3-TTS 12Hz 1.7B Base is an open-weights text-to-speech model from Alibaba (Qwen), released January 2026. Its published size is 1.7B parameters. Licence: apache-2.0.
Catalogue identity notes
Reference-audio voice-cloning Base checkpoint Qwen/Qwen3-TTS-12Hz-1.7B-Base. Distinct from the CustomVoice/VoiceDesign and 0.6B checkpoints. ZONOS2 'Qwen 3 TTS 1.7B' and Boson 'Qwen3-TTS-1.7B' are mapped here because they ran reference-audio cloning and this is the only released 1.7B cloning checkpoint.
Model information sources
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Qwen3-TTS 12Hz 1.7B Base, out of 148 with a score, then more from Alibaba (Qwen).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 23 |  Higgs Audio V3 TTS | 65±3 | api $15.00/M chars +local | ||
| 24 |  Raon-OpenTTS-0.3B | 65±2 | comparison only | ||
| 25 | Qwen3-TTS 12Hz 1.7B Base | 65±1 | local | ||
| 26 |  VoxCPM 0.5B | 64±1 | local | ||
| 27 | VoxCPM2 | 64±1 | local | ||
| more from Alibaba (Qwen) | |||||
| 44 | Qwen3-TTS 12Hz 0.6B Base | 62±2 | local | ||
| 98 | Qwen3-TTS 12Hz 0.6B CustomVoice | 46±6 | comparison only | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #11 of 37 · 23% of its score · 7 tests
Beats Gemini 3.1 Flash TTS Preview here, though they rank higher overall. Trails VoxCPM2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) +1 | 71.2% ✎ | █·█·██·▄·▄·▄ beats 7/10 | 8.2% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) | 76.9% ✎ | ▃··▃█▅▅··▃·▃ beats 6/10 | 4.3% |
| Seed-TTS test-en · Whisper large-v3 WER +1 | 1.6% ✎ | ▃··▃▅··▅·█▃▅ beats 11/11 | 3.5% |
| Seed-TTS-Eval · test-en/test-zh macro mean error | 1.3% ✎ | ▄··▄▄····█▄▄ beats 4/6 | 3.1% |
| Seed-TTS test-en WER (transcriber not stated) lab | 1.2% ✎ | ▄········█▄█ beats 5/5 | 1.9% |
| Seed-TTS test-zh · Paraformer CER | 0.9% ✎ | ▂··▄··▂▂█·▄▃ beats 11/11 | 1.6% |
| Seed-TTS test-zh error rate (transcriber not stated) lab | 0.8% ✎ | █··········█ beats 1/1 | — |
other tests22 not in a task · 77% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER +2 | 7.9% ✎ | █··██·██████ beats 5/8 | 13% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) +2 | 66.1% ✎ | ▄█▄·▄··▄··▄█ beats 1/8 | 8.7% |
| CV3-Eval hard-zh · CER +1 | 11.3% ✎ | ▃▃·▃··▃▃█··▃ beats 0/8 | 7.2% |
| CV3-Eval hard-en · DNSMOS (predicted quality) +1 | 3.87 / 5 ✎ | ▄·····█·▄▄·█ beats 4/6 | 6.9% |
| CV3-Eval EN (zero-shot English) · WER +1 | 3.8% ✎ | ▃·······▃▅█▅ beats 8/8 | 6.8% |
| CV3-Eval ZH (zero-shot Chinese) · CER | 3.1% ✎ | █▄···█▄·█··▄ beats 8/8 | 4.8% |
| ZTTS1-Eval in-the-wild · 10-language mean WER | 2.3% ✎ | ▄·····█··▄██ beats 7/7 | 4.1% |
| CV3-Eval multilingual voice clone · 9-language mean error lab +1 | 5.1% ✎ | ▄···▄··▄··██ beats 4/6 | 3.7% |
14 more tests
$ where to run it
Other ways to run it
version
- Run it as published · needs ~4.54 GB · from github.com/QwenLM
uv run --python 3.12 --with qwen-tts python -c "import torch, soundfile as sf; from qwen_tts import Qwen3TTSModel; model = Qwen3TTSModel.from_pretrained('Qwen/Qwen3-TTS-12Hz-1.7B-Base', device_map='cpu', dtype=torch.bfloat16); wavs, sr = model.generate_voice_clone(text='Hello, world.', language='English', 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.'); sf.write('output_voice_clone.wav', wavs[0], sr)"
Official weights and documentation: huggingface.co/Qwen ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 32%, Raon-OpenTTS: Open Models and… 27%, Qwen-Audio-3.0-TTS technical… 17% and 6 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 UltraEval-Audio (OpenBMB) · TTS replication results | 9 | 32% |
 Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech (arXiv:2605.20830v2) | 9 | 27% |
| Qwen-Audio-3.0-TTS technical report · lab-reported | 5 | 17% |
| ZONOS2 Technical Report (ZTTS1-Eval results) | 4 | 14% |
 Higgs Audio v3 comparison tables in SGLang-Omni cookbook (Boson AI contribution) | 3 | 4% |
| Qwen/Qwen3-TTS-12Hz-1.7B-Base model card · lab-reported | 1 | 2% |
| ZTTS1-Eval results snapshot (Zyphra) | 1 | 1% |
| RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2) | 2 | 1% |
| balacoon speech_gen_baselines · zero-shot TTS metrics (VCTK, DAPS-celeb) | 4 | 1% |
| prior · pull toward the middle | 0% | |
| weighted average | = 64.7 |
not counted (1)
| Qwen/Qwen3-TTS-12Hz-1.7B-Base model card | Seed-TTS test-zh error rate (transcriber not stated) [correct] | 0.8% | fewer than three models have this test yet, or this source ran it for this model alone |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json