ovos-tts-plugin-pocket-tts

OVOS TTS plugin for Pocket TTS by Kyutai — a lightweight, CPU-only text-to-speech engine with a 100M parameter model.

~6x real-time on Apple Silicon, no GPU required. Supports voice cloning from audio samples.

Install

pip install ovos-tts-plugin-pocket-tts

With uv, torch automatically resolves to the CPU-only wheel (configured via tool.uv.sources).

Linux: CPU-only torch (saves ~2GB)

On Linux, pip defaults to the CUDA torch wheel (~2.5GB). If you don't need GPU support (pocket-tts is CPU-only anyway), install torch from the CPU index first:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install ovos-tts-plugin-pocket-tts

On macOS, this is not needed — PyPI torch is already CPU-only (~60MB).

Configuration

{
  "tts": {
    "module": "ovos-tts-plugin-pocket-tts",
    "ovos-tts-plugin-pocket-tts": {
      "voice": "alba"
    }
  }
}

Voice options

Built-in voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma

Voice cloning: Set voice to a path to a .wav file, a pre-exported .safetensors file, or a HuggingFace URI (hf://kyutai/tts-voices/...).

Language support

This plugin auto-detects which languages your installed pocket-tts supports.

• Legacy pocket-tts (≤ 1.1.1, current PyPI release): English only.
• Multilingual pocket-tts (kyutai-labs/pocket-tts#155, not yet released): English, French, German, Spanish, Italian, Portuguese.

When the multilingual release lands on PyPI, just pip install -U pocket-tts and the plugin lights up the new languages automatically — no plugin upgrade required.

How language selection works

The plugin maps the active OVOS language (BCP-47, e.g. fr-FR) to a kyutai pocket-tts model id. Defaults:

Lookup tries the full BCP-47 tag first (e.g. pt-br), then falls back to the base subtag (pt), then to English. Unknown languages fall back to English with a log line.

Note: The _24l (24-layer) models are larger preview builds, not the final distilled releases. They are slower than the english defaults but support int8 quantization for a meaningful speedup. The plugin enables quantize=True automatically for any *_24l model. You can override per-language via the quantize config knob.

Override the language map

If you want, for example, the older english_v1 model, or to register a future regional model, set language_aliases in your config:

{
  "tts": {
    "module": "ovos-tts-plugin-pocket-tts",
    "ovos-tts-plugin-pocket-tts": {
      "voice": "alba",
      "language_aliases": {
        "en": "english_v1",
        "pt-br": "portuguese_brazil_24l"
      },
      "quantize": {
        "french_24l": false
      },
      "preload_languages": ["en", "fr"]
    }
  }
}

Memory note: Each loaded language model holds ~100M parameters in RAM (the *_24l previews are ~4× bigger before distillation). The plugin caches one model per used language, so leaving preload_languages empty and letting the cache warm on demand keeps the resident set small.

Roadmap

• SouraTTS (emotional TTS built on pocket-tts) is being tracked as a separate plugin (ovos-tts-plugin-soura-tts) — its emotion/intensity dimensions don't fit the current OVOS TTS interface and bundling would tie pocket-tts upgrades to SouraTTS releases.

License

Apache-2.0