MOSS-TTS

Why we included this project

MOSS-TTS is a speech and sound generation family aimed at real product use rather than demo-grade synthesis. It covers stable long-form narration, multi-speaker dialogue, and character voice design, and it goes beyond plain text-to-speech: short-text voice cloning, fine-grained pause and prosody control, and explicit language tags for code-switched synthesis are built in. Sound-effect generation is part of the same codebase, which also includes a roughly 100M-parameter model that streams on a handful of CPU cores, so the same stack can serve on-device and cloud workloads. Because it ships with inference backends like SGLang and vLLM plus a llama.cpp path for quantized weights on modest GPUs, teams can prototype quickly and later fine-tune on their own voice data without assembling the serving stack from scratch.