MOSS-TTS

Evaluate voice cloning and natural-language voice design locally through a unified speech endpoint using MOSS-TTS.

The default 4B clone model and 1.7B VoiceGenerator swap in GPU memory when the voice mode changes, so the first request after a switch takes longer. The 8B clone model requires more VRAM.

Setup

Run commands from the TomoriBot repository root using Python 3.12 and a driver compatible with CUDA 12.8:

Windows PowerShell

python -m venv servers\tts\moss\.venv
servers\tts\moss\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install --extra-index-url https://download.pytorch.org/whl/cu128 "moss-tts[torch-runtime] @ git+https://github.com/OpenMOSS/MOSS-TTS.git"
python -m pip install -r servers\tts\moss\requirements.txt
python servers\tts\moss\prefetch_models.py
python servers\tts\moss\server.py

Linux or WSL Bash

python3.12 -m venv servers/tts/moss/.venv
source servers/tts/moss/.venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --extra-index-url https://download.pytorch.org/whl/cu128 "moss-tts[torch-runtime] @ git+https://github.com/OpenMOSS/MOSS-TTS.git"
python -m pip install -r servers/tts/moss/requirements.txt
python servers/tts/moss/prefetch_models.py
python servers/tts/moss/server.py

The prefetch command downloads the clone model, VoiceGenerator, and audio tokenizers into your Hugging Face cache before starting the server. If disk space is limited, set HF_HOME to a larger partition. To download only one model, pass --mode clone or --mode voice-design.

The default endpoint is http://127.0.0.1:8018. Run bun run launch --moss to launch the server alongside TomoriBot. Auto mode prewarms the clone model from local cache. Set MOSS_TTS_WARM_MODE=voice-design to prewarm VoiceGenerator instead, or MOSS_TTS_WARM_MODE=none for lazy initialization. Check GET /health for active warm_mode, active_mode, and model_id. The wrapper uses Hugging Face trust_remote_code=True, so review upstream code before updates.

Register in TomoriBot

In /providers, choose Add New Custom Endpoint, set API Compatibility to tts-clone, and use endpoint URL http://127.0.0.1:8018. Add a Speech model with Voice Source Mode set to Auto and Script Markup set to Plain. Activate it under /config > Models > Switch Models.

For voice cloning, upload a clean reference clip under /config > Models > TTS Parameters & Voices and assign it under Persona > Voice. Shorter, clean audio clips yield the most consistent results. For voice design, save a natural-language description under Persona > Voice. Note that MOSS-VoiceGenerator is designed for English and Chinese. While the 4B clone model supports Japanese, explicit language tags improve synthesis clarity.

TomoriBot’s clone adapter does not send language tags automatically. For single-language use, set MOSS_TTS_DEFAULT_LANGUAGE=Japanese (or English, Chinese, etc.) before starting the server. Manual /synthesize calls can pass language directly.

The server reads its own shell environment; settings in the bot’s .env do not apply to an independently started Python terminal.

To run the 8B model on high-memory hardware, set MOSS_TTS_CLONE_MODEL_ID=OpenMOSS-Team/MOSS-TTS-v1.5 before prefetching. MOSS_TTS_PORT, MOSS_TTS_DEVICE, MOSS_TTS_DTYPE, and MOSS_TTS_MAX_NEW_TOKENS are configurable in .env.optional.example. Increase TTS_SYNTHESIZE_TIMEOUT_MS in TomoriBot if model swaps or CPU execution cause timeouts.