MOSS-TTS is an open-source family of speech models from Fudan University’s OpenMOSS team, built for long-form and dialogue speech synthesis. It goes beyond single-sentence text-to-speech: the family covers voice design from a text description, multi-speaker dialogue, sound-effect generation and real-time streaming TTS. The project has about 4,200 stars on GitHub and is actively developed in 2026.
The models are built on an audio-tokenizer architecture that represents speech as discrete tokens an LLM-style backbone can model, which is what enables coherent prosody over long passages and natural turn-taking in generated conversations. Released members include a flagship model for high-fidelity long-form narration, a voice-design model that creates new voices from descriptive prompts, a sound-effect model, and a lightweight streaming variant for interactive applications with low time-to-first-audio. The models are multilingual, support voice cloning from short reference clips, and ship with inference code and checkpoints under the Apache-2.0 license. OpenMOSS develops the family alongside its MOSS speech-understanding research, and the TTS line is aimed squarely at production use cases like audiobooks, podcasts, dubbing and conversational agents.
The family is developed by the OpenMOSS team at Fudan University, also known for the MOSS large language models, and is designed as a modular research-to-product stack: the shared audio tokenizer and backbone mean improvements in one member — say, better streaming latency — propagate to the others. Checkpoints, inference scripts and demos are published openly, and the dialogue model in particular targets a gap most open TTS ignores: generating two-voice conversations with natural pacing and speaker consistency rather than stitching single-speaker clips together after the fact.
MOSS-TTS is free and open source under the Apache-2.0 license — download the weights and self-host.
MOSS-TTS is for developers and creators building narration, dubbing, audiobook or voice-agent products who want an open, self-hostable speech engine.