MOSS-TTS: Generate Long-Form Speech, Dialogue, and Sound

Overview

MOSS-TTS is OpenMOSS’s family of speech and sound generation models, built around Qwen3-based language-model architectures and the MOSS-Audio-Tokenizer. It addresses cases where ordinary text-to-speech can struggle, including long, stable narration, voice cloning, multilingual and code-switched speech, multi-speaker dialogue, and speech generated incrementally for real-time interaction.

The family includes separate models for narration, spoken dialogue, voice design, real-time speech, and sound effects. You can use the Python code and model weights locally, connect through documented API and serving integrations, or choose a torch-free inference path for supported deployments. The repository is most relevant to teams building audio products, rather than people looking for a single small TTS package.

Key Features

• Generate long-form speech with voice cloning and controls for pronunciation, duration, and pauses.
• Choose between MossTTSDelay for long-context stability and production use, MossTTSLocal for a local-transformer approach, and MossTTSRealtime for conversational streaming.
• Create multi-speaker dialogue with MOSS-TTSD, or design voices from text descriptions with MOSS-VoiceGenerator.
• Generate environmental and other sound effects with the MOSS-SoundEffect models.
• Use v1.5 releases for 31-language support, language tags, more stable cloning, and explicit pause markers; the local-transformer v1.5 model supports native 48 kHz stereo audio.
• Run inference through Python and Transformers, or use documented integrations with SGLang-Omni and vLLM-Omni. A llama.cpp and ONNX Runtime path is available for torch-free inference.

Use Cases

• Audiobook and narration teams: synthesize longer scripts where stable speaker identity, pronunciation, and pacing matter.
• Podcast and dubbing developers: generate multi-speaker dialogue with MOSS-TTSD rather than stitching together separate single-speaker clips.
• Voice-agent builders: use MOSS-TTS-Realtime when an application needs incremental speech output and multi-turn voice context.
• Audio and game creators: create prompted voices or environmental sound effects without recording every asset.
• ML engineers: fine-tune model architectures or deploy inference through the supported serving backends.

Getting Started

The README recommends Python 3.12 and provides Conda and uv setup paths. For the Conda path:

git clone https://github.com/OpenMOSS/MOSS-TTS.git
cd MOSS-TTS
pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[torch-runtime]"

Then follow the README’s model-specific examples and select the checkpoint that matches your task. The default PyTorch setup uses CUDA wheels; the repository documents alternatives, including CPU installation and torch-free inference.

Alternatives

• chatterbox-vllm: Chatterbox is a vLLM-optimized TTS port focused on batched generation and voice conditioning, rather than MOSS-TTS’s broader speech and sound generation.
• Spark-TTS: Spark-TTS uses Qwen2.5 for Chinese and English synthesis with zero-shot cloning, while MOSS-TTS uses Qwen3-based models and also generates sound effects.
• GPT-SoVITS: GPT-SoVITS centers on voice cloning, fine-tuning, and WebUI workflows, rather than MOSS-TTS’s broader generation features and streaming.
• ChatTTS: ChatTTS focuses on Chinese and English dialogue speech with prosody controls, while MOSS-TTS also supports narration, voice design, and sound effects.

Considerations

• Checkpoint sizes range from about 1.3B to 8B parameters, so hardware needs vary substantially by model. The README also documents a CPU-first Nano model in a separate repository; do not assume the main family’s larger checkpoints are suitable for CPU-only use.
• The standard setup uses Python and PyTorch; FFmpeg is required for audio I/O through torchcodec. FlashAttention is optional and only applies to supported GPUs. The separate llama.cpp route avoids PyTorch but has its own dependencies and setup steps.
• The project is marked maintained, with recent commits, but has no GitHub releases in the past year. Teams needing versioned releases should account for that in their update and deployment process.
• The README describes several architectures and model-specific capabilities. Confirm language coverage, audio format, and serving-backend support for the exact checkpoint you plan to deploy.
• The repository and its models are licensed under Apache-2.0. Review the license terms when redistributing or incorporating them into a product.