One repo, five models: an open-source speech family

What it does

MOSS-TTS is a collection of five production-ready audio models that cover the full speech pipeline. The flagship MOSS-TTS handles long-form narration and zero-shot voice cloning with phoneme-level control; MOSS-TTSD generates multi-speaker dialogues; MOSS-VoiceGenerator designs new voices from text prompts alone; MOSS-TTS-Realtime targets low-latency voice agents with incremental synthesis; and MOSS-SoundEffect generates environmental audio and soundscapes. A separate MOSS-TTS-Nano variant squeezes multilingual streaming onto four CPU cores.

The interesting bit

The project treats TTS as infrastructure, not a demo. It exposes two distinct architectures—MossTTSDelay for stable long-context work and MossTTSLocal for lightweight streaming—and can run entirely without PyTorch via llama.cpp and ONNX, allowing an 8B model to fit on an 8GB GPU.

Key highlights

• Five specialized models instead of a single generalist: TTS, dialogue (TTSD), voice design (VoiceGenerator), real-time streaming (Realtime), and sound effects (SoundEffect).
• MOSS-TTS-Realtime claims a 180 ms time-to-first-byte and incremental synthesis for multi-turn voice agents.
• MOSS-TTS-Nano runs ~100M parameters and streams 48 kHz stereo audio on four CPU cores.
• A PyTorch-free inference path using llama.cpp and ONNX Runtime, with an SGLang backend that reportedly pushes MossTTSDelay throughput up by roughly 3×.
• Fine-grained control over duration, Pinyin, phonemes, and explicit pauses via markup like [pause X.Ys] in the flagship model.

Verdict

Developers building voice agents, audiobook pipelines, or content-generation tools should look here—especially if you need to deploy offline or on modest hardware. If you just need a quick, single-speaker TTS snippet and don’t care about latency or long-form stability, the complexity of a five-model family might be overkill.