MOSS-TTS
#28 en Synthèse vocale (TTS)Moss · depuis 10. Februar 2026 (MOSS-TTS Family); MOSS-TTS-Nano: 10. April 2026 · 3× · vu le 29 juin 2026
1
Momentum
MOSS-TTS is an open-source family of speech and sound generation models from MOSI.AI and the OpenMOSS team (Shanghai Innovation Institution, Fudan NLP Lab). The family includes several specialized models for high-fidelity long-form narration, multi-speaker dialogue, voice design, environmental sound effects, and real-time streaming TTS. It also includes MOSS-TTS-Nano, a very small model (0.1B parameters) for CPU-based real-time speech generation with 48kHz stereo output in up to 20 languages. All models are released under the Apache 2.0 license and are approved for commercial use.
Historique du momentum
19.05.17.08.
Fonctionnalités
| Real-Time Streaming | Yes, streaming output with low first-token latency, including automatic chunking of long text |
| Latency | MOSS-TTS-Realtime: ~180ms first-byte latency; Nano: real-time factor < 1.0 on 4 CPU cores |
| License | Apache License 2.0 |
| Platform | Open-source on GitHub/Hugging Face; runs locally (GPU for flagship 8B, CPU for Nano); supported by vLLM-Omni, SGLang, ComfyUI, mlx-audio, ONNX, llama.cpp |
| Price | Free, open-source model weights (Apache 2.0) for self-hosting |
| Release Date | MOSS-TTS Family: February 10, 2026; MOSS-TTS-Nano: April 10, 2026 |
| Languages | 20 languages (incl. Chinese, English, German, Spanish, French, Japanese, Korean, Arabic, Persian) |
| Voice Cloning | Zero-shot voice cloning from short reference audio (3-15 seconds), no fine-tuning required |