Speech generation · Opt-in
MOSS-TTS v1.5
An 8B flagship: 31 languages, zero-shot cloning and duration control. Wants a 16 GB+ GPU.
- OpenMOSSBy
- 31Languages
- 24 kHzOutput
- ~16 GBDownload
Can do
- Voice cloningYes. YesSpeaks in a voice copied from a short reference clip.
- Voice designNo. NoBuilds a new voice from a written description.
- Emotion controlNo. NoTakes graded emotion from a clip, a vector or text.
- Preset voicesNo. NoShips fixed voices; no reference clip needed.
Languages
31languages
Full list on the model card Runs on
PyTorch, isolated
isolatedPyTorchCUDAROCmIntel Arc · experimentalNPU · experimentalCPUEngine doc
The bf16 weights are about 16 GB, so a 16 GB+ GPU is the realistic CUDA target.
Hardware
| Engine | CUDA | ROCm | Intel Arc | Apple Silicon | Vulkan | NPU | CPU |
|---|---|---|---|---|---|---|---|
| PyTorch, isolated | Supported | Supported | Experimental | Not supported | Not supported | Experimental | Supported |
Supported Needs a build Experimental
Licence
- Code
- Apache-2.0
- Weights
- Apache-2.0
- XPU and NPU selection is covered by mocked tests; synthesis is not certified on that hardware.
VoiceStudio is AGPL-3.0 and doesn't relicense model weights. Licence guide
Related
From the VoiceStudio app at 0834c8b. 27 models listed.