OpenMOSS/MOSS-TTS

An open-source model family for long-form speech, dialogue synthesis, voice design, sound effects, and real-time streaming TTS

What it solves

MOSS-TTS Family is a collection of open-source speech and sound generation models designed to handle complex, real-world audio production. It addresses the limitations of single TTS models by providing specialized tools for high-fidelity long-form speech, expressive multi-speaker dialogues, zero-shot voice design from text prompts, low-latency real-time interaction, and environmental sound effect generation.

How it works

The project utilizes three primary architectures to balance stability, speed, and latency:

  • MossTTSDelay: Uses multi-head parallel RVQ prediction with delay-pattern scheduling, optimized for long-context stability and production readiness.
  • MossTTSLocal: Employs time-synchronous RVQ blocks with a depth transformer, focusing on lightweight flexibility and streaming performance.
  • MossTTSRealtime: A capability-driven design that models multi-turn context from both text and user acoustics to enable low-latency streaming speech for voice agents.

Additionally, the family includes a DiT backbone with Flow Matching for the MOSS-SoundEffect-v2.0 model to generate high-quality bilingual sound effects.

Who it’s for

This family of models is intended for developers and creators building:

  • Content Creators: For long-form narration, podcasts, and dubbing.
  • AI Voice Agents: For real-time, low-latency interactive voice systems.
  • Game and Film Developers: For generating environmental sound effects and diverse character voices.
  • Researchers: For those needing reproducible references for speech synthesis and audio generation.

Highlights

  • Diverse Model Suite: Includes specialized models for general TTS, dialogue (MOSS-TTSD), voice design (MOSS-VoiceGenerator), real-time agents (MOSS-TTS-Realtime), and sound effects (MOSS-SoundEffect).
  • High Fidelity: Supports 48 kHz stereo output via the MOSS-Audio-Tokenizer-v2.
  • Extreme Flexibility: Offers a range of model sizes from ~100M (Nano) to 8B parameters.
  • Broad Language Support: MOSS-TTS-v1.5 supports 31 different languages.
  • Deployment Ready: Supports torch-free inference via llama.cpp and ONNX Runtime, and is integrated with accelerated backends like SGLang-Omni and vLLM-Omni.

Related

  • Project
    OpenMOSS/MOSS-TTSD

    A long-form spoken dialogue generation model that transforms scripts into expressive multi-party conversations with support for up to 60 minutes of continuous audio.

  • Project
    2noise/ChatTTS

    ChatTTS is a generative text-to-speech model optimized for conversational dialogue, providing natural, expressive speech with control over laughter and pauses.

  • Project
    resemble-ai/chatterbox

    A family of open-source text-to-speech models providing low-latency voice cloning and multilingual speech generation across various model sizes for GPU and CPU deployment.

  • Project
    rsxdalv/TTS-WebUI

    A unified web interface and manager for a wide variety of open-source text-to-speech, audio generation, and audio conversion AI models.

  • Project
    OpenMOSS/MOSS-TTS-Nano

    A multilingual, tiny speech generation model with 0.1B parameters designed for real-time, CPU-based inference and high-fidelity voice cloning.