Transformer Pipelines

August 1, 2026 · View on GitHub

The server can run OVOS transformer plugins around transcription. Both hooks live in the model containers, so every surface gets them: the native /stt endpoint, all vendor-compat routers, the websocket streaming routes, MCP and UTCP.

  • Audio transformers process the audio before the STT stage. When the chain includes an AudioLanguageDetector, its detected language resolves lang=auto (an explicitly requested language always wins).
  • Utterance transformers rewrite the transcript before it is returned.

Configuration

Loading is config-gated and opt-in via the standard mycroft.conf sections; with no config the server behaves exactly as before:

{
  "audio_transformers": {
    "ovos-audio-transformer-plugin-speechbrain-langdetect": {}
  },
  "utterance_transformers": {
    "ovos-utterance-corrections-plugin": {}
  }
}

Chains run in ascending priority order (OVOS-TRANSFORM §4); an explicit "order" list in a section wins over priorities. In multi-model mode (--multi) one set of transformer instances is shared across the per-language engines. See the ovos-plugin-manager transformer docs for the full contract.

When to use — and the surprise factor

An utterance transformer on the server means clients receive a different transcript than what the STT engine heard — a client sends audio saying "who is the speaker of teh house" and gets back a corrected string it can't distinguish from raw STT output. Do it on purpose:

  • Fleet-wide corrections: domain vocabulary fixes (product names, callsigns, jargon) applied once, for every client of this server.
  • Language routing: an AudioLanguageDetector in the audio chain lets heterogeneous clients send lang=auto and still hit the right model.
  • Audio cleanup: server-side denoise before STT when clients are thin and can't do it themselves.

If the effect should be per-device (one bad microphone needs denoise), run the audio transformer on that device instead — and never enable the same plugin on both sides, or it is applied twice. Similarly, if the OVOS stack consuming this server already runs utterance transformers (ovos-core does), enable each text plugin in exactly one of the two places.

← API compatibility · Home · Wyoming integration →