#3030th of 83 in Open modelsFor voice, speech and music
moss-tts
Open-weight TTS model family for cloning, dialogue, realtime speech and sound effects
- Stars
- 4.3k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
MOSS-TTS is a family of open-weight speech and sound generation models from MOSI.AI and OpenMOSS, built on the MossTTSDelay, MossTTSLocal and MossTTSRealtime architectures. It covers long-form multilingual TTS with voice cloning (31 languages in v1.5), multi-speaker dialogue, voice design from text prompts, realtime synthesis at 180 ms TTFB, and sound effects. The repo has inference code, fine-tuning tutorials, and llama.cpp/ONNX, SGLang-Omni and vLLM-Omni backends.
Who it is for: Engineers self-hosting TTS, voice cloning or voice agents with open weights
Strengths
- Covers cloning, dialogue, voice design, realtime and sound effects, from ~100M to 8B parameters
- Torch-free path via llama.cpp GGUF weights plus ONNX codec; 8B fits on 8 GB GPUs
- Serving via SGLang-Omni (OpenAI-compatible /v1/audio/speech), vLLM-Omni and mlx-audio
- Fine-tuning tutorials for the Delay, Local and Realtime architectures
Weaknesses
- Separate checkpoints on different architectures; no single unified entry point
- 8B flagship models need a sizable GPU; CPU use means the separate MOSS-TTS-Nano
- No release tags and no Docker or compose files in the repo
- Several components live in other repos (TTSD, Nano, llama.cpp fork, tokenizer)
What it needs
- GPU optional
- Needs PyTorch, Hugging Face Transformers, FlashAttention 2 (optional), llama.cpp (optional), ONNX Runtime (optional)
- Models: MOSS-TTS-v1.5 (8B), MOSS-TTS-Local-Transformer-v1.5 (4B), MOSS-TTSD-v1.0, MOSS-VoiceGenerator, MOSS-TTS-Realtime
Also in Open models
See all 83| Rank | Project | Score |
|---|---|---|
| 1 | ruviewWiFi CSI sensing on ESP32 for presence, breathing, heart rate and pose | 94 out of 100 |
| 2 | layaSingle-pass multilingual classifier returning typed decisions with calibrated odds | 88 out of 100 |
| 3 | rf-detrReal-time transformer models for object detection, segmentation and keypoints | 82 out of 100 |
| #4 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 78 out of 100 |
| #5 | open_clipPyTorch library for training and running CLIP-style image-text models | 78 out of 100 |