MOSS-TTS — speech and sound generation models

★ 4.1K

MOSS-TTS Family is an open-source family of speech and sound generation models from MOSI.AI and the OpenMOSS team, covering long-form narration, multi-speaker dialogue, voice cloning and voice design, sound effects and real-time streaming speech. Reach for it when you need to voice a long text or a podcast with several speakers; when you need a reference voice for dubbing; when you need low-latency streaming speech; or when you need ambient sounds for a game or video. The family includes MOSS-TTS-v1.5 (multilingual narration, stable voice cloning, pause control), MOSS-TTSD (dialogue, podcasts, dubbing), MOSS-TTS-Realtime (streaming speech), MOSS-TTS-Nano (about 100M parameters, runs on 4 CPU cores), MOSS-SoundEffect-v2 (text to audio up to 30 seconds, DiT with Flow Matching) and Local-Transformer-v1.5, a 4B model with 48 kHz stereo output via MOSS-Audio-Tokenizer-v2. Inference works through vLLM-Omni and SGLang-Omni with an OpenAI-compatible /v1/audio/speech endpoint, and through llama.cpp with GGUF and mlx-audio. Weights are on Hugging Face, the code is Python and the licence is Apache 2.0. The direction is generating speech and sound, not recognising it. Clone a voice only with its owner's consent. For speech-to-text you need a different tool.