whatmodel.sh / tts / MOSS-TTS v1.5 (8B)
MOSS-TTS v1.5 (8B)
↗ share
+ contribute
MOSS-TTS v1.5 (8B) is an open-weights text-to-speech model from OpenMOSS, released May 2026. Its published size is 8.49B parameters. Licence: apache-2.0.
Catalogue identity notes
OpenMOSS-Team/MOSS-TTS-v1.5 (MossTTSDelay-8B; 8,489,841,664 parameters per HF safetensors metadata). Distinct from MOSS-TTSD dialogue models.
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 MOSS-TTS v1.5 (8B), out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 31 |  Fish Audio S2.1 Pro | 63±3 | api $15.00/M chars |
| 32 |  IndexTTS-2 | 63±1 | local |
| 33 | MOSS-TTS v1.5 (8B) | 63±2 | local |
| 34 |  GLM-TTS | 63±2 | local |
| 35 |  Sonic 3.5 | 63±2 | comparison only |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #22 of 37 · 45% of its score · 6 tests
Beats Gemini 3.1 Flash TTS Preview here, though they rank higher overall. Trails E2 TTS Base (F5-TTS reproduction), DMOSpeech 2, Qwen3-TTS 12Hz 0.6B Base here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) | 68.0% ✎ | █·█·██·▄·▄·▄ beats 3/10 | 11% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) | 76.1% ✎ | ▃··▃█▅▅··▃·▃ beats 4/10 | 11% |
| Seed-TTS-Eval · test-en/test-zh macro mean error | 1.7% ✎ | ▄··▄▄····█▄▄ beats 1/6 | 7.5% |
| MiniMax multilingual test set · 21-language mean error (UltraEval-Audio common set) | 5.9% ✎ | █··█·····█·█ beats 2/3 | 6.2% |
| Seed-TTS test-en · Whisper large-v3 WER +1 | 2.7% ✎ | ▃··▃▅··▅·█▃▅ beats 1/11 | 4.8% |
| Seed-TTS test-zh · Paraformer CER +1 | 1.5% ✎ | ▃··▄··▂▂█·▄▂ beats 0/11 | 3.8% |
other tests9 not in a task · 54% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER | 6.4% ✎ | █··██·██████ beats 8/8 | 15% |
| CV3-Eval ZH (zero-shot Chinese) · CER | 3.9% ✎ | █▄···█▄·█··▄ beats 2/8 | 12% |
| CV3-Eval hard-zh · CER | 8.4% ✎ | ▃▃·▃··▃▃█··▃ beats 5/8 | 9.4% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) | 66.4% ✎ | ▄█▄·▄··▄··▄█ beats 2/8 | 8.8% |
| CV3-Eval EN (zero-shot English) · WER | 5.5% ✎ | ▃·······▃▅█▅ beats 4/8 | 3.8% |
| RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity | 57.0% ✎ | ▃·····▂▂▂▃█▆ beats 10/15 | 3.2% |
| MiniMax multilingual test set · 23-language mean error | 3.8% ✎ | ▃·▃·▃·····▃█ beats 4/6 | 1.6% |
| CV3-Eval multilingual voice clone · 9-language mean error | 6.1% ✎ | ▃·····▃···▅█ beats 3/6 | 1.2% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| RVCBench LibriTTS clean voice cloning · Whisper-medium WER | 6.0% ✎ | ▂▂▂··▂·▃▃▃█▂ beats 9/15 | 0.0% |
$ where to run it
Other ways to run it
version
- Run it as published · needs ~24.08 GB · from huggingface.co/OpenMOSS-Team (official)
conda create -n moss-tts python=3.12 -y && conda activate moss-tts && git clone https://github.com/OpenMOSS/MOSS-TTS.git && cd MOSS-TTS && pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e . && python -c "import torch, torchaudio; from transformers import AutoModel, AutoProcessor; torch.backends.cuda.enable_cudnn_sdp(False); pretrained_model_name_or_path = 'OpenMOSS-Team/MOSS-TTS-v1.5'; device = 'cuda' if torch.cuda.is_available() else 'cpu'; dtype = torch.bfloat16 if device == 'cuda' else torch.float32; attn_implementation = 'sdpa' if device == 'cuda' else 'eager'; processor = AutoProcessor.from_pretrained(pretrained_model_name_or_path, trust_remote_code=True); processor.audio_tokenizer = processor.audio_tokenizer.to(device); model = AutoModel.from_pretrained(pretrained_model_name_or_path, trust_remote_code=True, attn_implementation=attn_implementation, torch_dtype=dtype).to(device); model.eval(); torch.set_grad_enabled(False); batch = processor([[processor.build_user_message(text='We stand on the threshold of the AI era.')]], mode='generation'); outputs = model.generate(input_ids=batch['input_ids'].to(device), attention_mask=batch['attention_mask'].to(device), max_new_tokens=4096); [torchaudio.save('sample{}.wav'.format(i), m.audio_codes_list[0].unsqueeze(0), processor.model_config.sampling_rate) for i, m in enumerate(processor.decode(outputs))]"
Official weights and documentation: huggingface.co/OpenMOSS-Team ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 85%, Higgs Audio v3 comparison tab… 10%, RVCBench: Benchmarking Robust… 3% and 1 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 | 10 | 85% |
 Higgs Audio v3 comparison tables in SGLang-Omni cookbook (Boson AI contribution) | 3 | 10% |
 RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2) | 2 | 3% |
| SGLang-Omni serving cookbooks (SGLang-Omni team measurements) | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 63.2 |
not counted (1)
| SGLang-Omni serving cookbooks (SGLang-Omni team measurements) | Seed-TTS test-zh · Paraformer CER [correct] | 1.4% | the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts |
$ 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