ovos-tts-plugin-omnivoice

TTS plugin for OpenVoiceOS wrapping OmniVoice, a massively-multilingual (600+ language) zero-shot text-to-speech model from the k2-fsa / sherpa / icefall team, built on a diffusion language-model architecture.

OmniVoice has no fixed speaker catalogue. It produces a voice in one of three modes, all wired through this plugin:

Languages

OmniVoice supports 600+ languages. This plugin ships explicit BCP-47 → model language_id mappings for the languages below. Any other tag falls back to its primary sub-tag, which the plugin passes straight to the model, so unmapped languages still work if the model knows them.

Set the OVOS lang (or per-request lang) to any of these tags. To reach one of the 600+ languages that isn't mapped here, pass its ISO code as the lang. The model accepts a language name ("English") or code ("en") directly.

Arabic varieties

OmniVoice was trained on many Arabic varieties, so regional Arabic tags route to a distinct trained variety instead of collapsing to Modern Standard Arabic:

Voice design across languages: OmniVoice's voice-design (instruct) mode is trained mainly on English and Chinese. It generalizes to other languages but is less stable than the auto voice or a cloned reference clip. For the most stable output in other languages, prefer voice: "auto" or supply ref_audio.

Install

The OmniVoice runtime (PyTorch + the multi-GB omnivoice package) is a core dependency, so a plain install is enough to synthesize:

pip install ovos-tts-plugin-omnivoice

It runs on CPU or GPU. A GPU is optional and only makes synthesis faster; it is not required. The default install pulls PyPI's default torch build (CPU-capable). To target a specific accelerator, reinstall torch from the matching index:

# NVIDIA CUDA:
pip install torch==2.8.0+cu128 torchaudio==2.8.0+cu128 \
    --extra-index-url https://download.pytorch.org/whl/cu128
# ...or AMD ROCm:
pip install torch torchaudio --index-url https://download.pytorch.org/whl/rocm6.2
# ...or a slim CPU-only build:
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu

The plugin auto-detects the device (CUDA/ROCm if available, else CPU). Weights download automatically from k2-fsa/OmniVoice on the first synthesis.

Configuration

mycroft.conf:

{
  "tts": {
    "module": "ovos-tts-plugin-omnivoice",
    "ovos-tts-plugin-omnivoice": {
      "lang": "en",
      "voice": "auto",
      "num_step": 32
    }
  }
}

Config keys:

Example: clone a reference voice:

{
  "tts": {
    "module": "ovos-tts-plugin-omnivoice",
    "ovos-tts-plugin-omnivoice": {
      "lang": "en",
      "ref_audio": "/home/ovos/voices/ref.wav",
      "ref_text": "Welcome to the voice assistant."
    }
  }
}

Serve it via ovos-tts-server

Install the server alongside this plugin, then point it at the module:

pip install ovos-tts-server ovos-tts-plugin-omnivoice

ovos-tts-server \
    --engine ovos-tts-plugin-omnivoice \
    --port 9666 \
    --cache

Then request audio (the lang selects the language / variety):

# English
curl -G "http://localhost:9666/synthesize/hello there" --data-urlencode "lang=en" -o en.wav

# Modern Standard Arabic
curl -G "http://localhost:9666/synthesize/مرحبا" --data-urlencode "lang=ar" -o msa.wav

# Najdi (Saudi) Arabic
curl -G "http://localhost:9666/synthesize/مرحبا" --data-urlencode "lang=ar-SA" -o najdi.wav

Docker

A batteries-included image runs the plugin as an ovos-tts-server:

docker run -p 9666:9666 -v omnivoice-cache:/home/ovos/.cache \
  ghcr.io/openvoiceos/ovos-tts-plugin-omnivoice:latest

The image is built and pushed to GHCR on every push to dev/master. See docs/docker.md for configuration and the bundled docker-compose.yml.

License

Apache-2.0. The k2-fsa team distributes OmniVoice itself under its own license. See the upstream repository.