MOSS-TTS: Technical Breakthroughs and Multi-Scenario Applications of the Open-Source Speech Synthesis Family

MOSS-TTS is an open-source speech synthesis model family co-launched by the OpenMOSS team and MOSI.AI, covering full-scenario needs such as high-quality long-text generation, multi-speaker dialogue, voice role design, and real-time streaming TTS. This family includes five core production-grade models with a modular design that can be used independently or in combination. Its technical indicators lead the open-source community, providing a complete toolchain from research to production and supporting full-stack deployment from cloud to edge.

Speech synthesis technology has moved from the laboratory to practical applications, but a single model can hardly meet complex requirements such as human-like realism, accurate pronunciation, style switching, stable long-text processing, and dialogue role-playing. Thus, the MOSS-TTS family was born as an open-source solution for real-world scenarios, breaking down the speech synthesis workflow into five combinable production-grade models and redefining the capability boundaries of open-source speech synthesis.

Five Core Models

1. MOSS-TTS: Flagship model focusing on high-fidelity zero-shot cloning, supporting long texts and multiple languages. Its 8B parameter architecture outperforms all open-source models on the Seed-TTS-eval benchmark.
2. MOSS-TTSD: Dialogue expert suitable for multi-speaker ultra-long dialogues, with subjective evaluation surpassing closed-source models like Doubao and Gemini 2.5-pro.
3. MOSS-VoiceGenerator: Open-source voice design model that generates diverse voices from text, with performance exceeding top peer models.
4. MOSS-TTS-Realtime: Real-time voice agent engine with a TTFB of only 180ms and end-to-end response of 377ms.
5. MOSS-SoundEffect: Sound effect generation model covering multiple audio categories, suitable for film and game applications.

Technical Architecture

• MossTTSDelay: Emphasizes long-context stability and production readiness, adopted by the 8B parameter model.
• MossTTSLocal: Lightweight and flexible, adopted by the 1.7B parameter model.
• MossTTSRealtime: Multi-turn context-aware with low-latency streaming output.

Audio Tokenizer

MOSS-Audio-Tokenizer is based on the Cat architecture with 1.6 billion parameters, supporting extreme compression (12.5Hz frame rate), large-scale general audio training, and native streaming design.

1. Standard PyTorch Deployment: Python 3.12+Transformers 5.0.0+CUDA 12.8, supporting FlashAttention2 and providing Gradio demos.
2. llama.cpp Torch-Free Inference: Lightweight edge deployment without PyTorch, allowing the 8B model to run with 8GB VRAM.
3. SGLang Acceleration: 3x throughput improvement, supporting integrated deployment of models and tokenizers.
4. MOSS-TTS-Nano: 100 million parameter CPU-first solution, enabling streaming generation on 4-core CPUs and supporting multi-language cloning.

• On Seed-TTS-eval, MossTTSDelay (8B) and MossTTSLocal (1.7B) rank first among open-source models in WER, CER, and SIM metrics.
• MOSS-TTSD-v1.0 leads in objective indicators, and its subjective evaluation surpasses closed-source models like ElevenLabs V3 and Gemini 2.5-pro.
• After warm-up, MOSS-TTS-Realtime achieves a TTFB of 180ms, a real-time factor of 0.51, and an end-to-end first-sentence response of 377ms.

• Language Support: Covers 20 languages (Chinese, English, German, French, etc.).
• Ecosystem Integration: Entered the OpenClaw skill market, with community contributions including ComfyUI extensions and OpenAI-compatible APIs.
• Future: MOSS-TTS 2.0 will be released soon; the team will continue to iterate features and build an open, co-constructed ecosystem.