OVOS Transcription Validator Plugin

A plugin for OVOS that uses an OpenAI-compatible Large Language Model (LLM) to validate transcriptions from speech-to-text (STT) before they are processed by your voice assistant.

It helps filter out garbled, nonsensical, or incomplete utterances—reducing confusion and improving the accuracy of downstream skills.

Features

• Multilingual transcription validation
• Powered by OpenAI-compatible LLMs (e.g., OpenAI API, Ollama, custom servers)
• Filters out invalid utterances before processing
• Optional feedback via error sound or dialog
• Fully configurable, with per-request overrides

How It Works

1. The plugin receives an STT transcription and language code.
2. A structured prompt with examples is sent to the configured OpenAI-compatible LLM API.
3. The LLM responds with True (valid) or False (invalid).
4. If invalid:
   • The utterance is canceled.
   • Optionally, a dialog prompt or error sound is triggered.

Installation

pip install ovos-transcription-validator-plugin

Configuration

Add the plugin to the utterance_transformers section of your mycroft.conf.

The plugin defaults to using https://llama.smartgic.io/v1 with qwen2.5:7b and a placeholder API key sk-xxxx.

{
  "utterance_transformers": {
    "ovos-transcription-validator-plugin": {
      "api_url": "https://llama.smartgic.io/v1",
      "api_key": "sk-xxxx",
      "model": "qwen2.5:7b",
      "prompt_template_system": "/path/to/system_template.txt",
      "prompt_template_user": "/path/to/user_template.txt",
      "error_sound": true,
      "mode": "ignore"
    }
  }
}

Available Settings

Requirements & Notes

• Requires an OpenAI-compatible LLM API endpoint to be accessible ( e.g., OpenAI API, Ollama running locally, or another custom server).
• You must have a supported model already available on your chosen LLM server.
• The plugin can adapt to different languages based on the LLM's capabilities and training.

Feedback & Contributions

Found a bug or want to contribute? PRs and issues are welcome!