MOSS-TTS
Open speech and sound model family for TTS, dialogue, and effects
What it is
MOSS-TTS is a set of open-weight models for generating speech and audio: long-form narration, multi-speaker dialogue, voice cloning, voice design from text prompts, environmental sound effects, and low-latency streaming speech. Sizes range from a roughly 100M-parameter model that runs on a CPU to a 4B model with 48 kHz stereo output, and the repo includes inference code, fine-tuning tutorials, and support for serving backends such as SGLang, vLLM-Omni, llama.cpp, and mlx-audio. Developers building podcasts, dubbing, narration, or voice features into their own products can run and adapt the models locally.