MOSS-TTS

Use servers/tts/moss/server.py to try MOSS voice cloning and text-described voice design through one local endpoint. Auto mode selects the clone model when TomoriBot sends ref_audio and MOSS-VoiceGenerator when it sends instruct. It keeps only one model loaded at a time. This is a trial sidecar, not a streaming Discord voice-chat integration.

The default clone model is MOSS-TTS-Local-Transformer-v1.5 (4B), chosen as the practical starting point for a 16 GB GPU. MOSS-TTS-v1.5 is an 8B alternative but will generally need more than 16 GB VRAM at BF16. Voice design uses MOSS-VoiceGenerator (about 1.7B). Auto mode swaps models rather than keeping both in VRAM, so a mode change still incurs a GPU load delay.

Setup

Run from the TomoriBot repository root. Use Python 3.12 and a CUDA driver compatible with the upstream CUDA 12.8 PyTorch wheels. Upstream’s runtime extra pins PyTorch and Torchaudio 2.9.1+cu128; keep this sidecar in its own virtual environment. Other CUDA or CPU stacks need a separately validated installation.

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 each model’s audio tokenizer into the Hugging Face cache before the server starts. It checks available cache-volume disk space before each repository download and reuses cached files, but both models need substantial space. If the check fails, free space or set HF_HOME to a larger volume in the shell before prefetching and starting the server. Run prefetch again after changing either model ID. To download only one mode for a limited trial, pass --mode clone or --mode voice-design; the other mode may still download on first use.

The endpoint is http://127.0.0.1:8018. Auto mode warms the clone model from the local cache before reporting startup complete. If the clone was not prefetched, startup fails rather than downloading it unexpectedly. MOSS_TTS_WARM_MODE=voice-design warms VoiceGenerator instead; MOSS_TTS_WARM_MODE=none keeps the previous lazy startup. Only one mode stays in GPU memory. Check GET /health for warm_mode, active_mode, and model_id. The wrapper uses Hugging Face trust_remote_code=True, so install only from a source you trust and review upstream changes before updating.

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 Auto and Script Markup Plain. Then activate it under /config > Models > Switch Models.

For cloning, upload a clean reference clip under /config > Models > TTS Parameters & Voices and assign it under Persona > Voice. For voice design, save a natural-language voice description under Persona > Voice instead. MOSS-TTS uses the audio reference; it does not use TomoriBot’s optional reference transcript. MOSS-VoiceGenerator is documented for English and Chinese, not Japanese. The 4B clone model supports Japanese, but a known language tag improves multilingual synthesis.

TomoriBot’s current clone adapter sends no language tag. For a single-language trial, set MOSS_TTS_DEFAULT_LANGUAGE=Japanese (or English, Chinese, etc.) before starting the server. A manual /synthesize request can instead supply language per request. Leave the variable unset for mixed-language use; evaluate Japanese output before relying on it.

The sidecar reads its own process environment. Adding a value to the bot’s .env does not automatically pass it to a separately started Python process.

To try the 8B flagship on a machine with enough memory, set MOSS_TTS_CLONE_MODEL_ID=OpenMOSS-Team/MOSS-TTS-v1.5 before prefetching. TOMORI_TTS_PORT, MOSS_TTS_DEVICE, MOSS_TTS_DTYPE, MOSS_TTS_MAX_REF_AUDIO_BYTES, and MOSS_TTS_MAX_NEW_TOKENS are also configurable in .env.optional.example. The bot’s TTS_SYNTHESIZE_TIMEOUT_MS may need increasing for mode swaps or CPU inference.