Audio Meeting Transcript

Problem

Meeting recordings often contain multiple speakers, making plain transcripts hard to audit: it is not always clear who said what, when it happened, or which decisions and action items emerged. This skill targets local audio processing. It first normalizes mp3, m4a, flac, and other files into 16kHz mono WAV, then uses offline Vosk for speech recognition, reducing the need to upload audio to external servers.

How It Works

The workflow has four stages: audio preprocessing, transcription and speaker separation, Markdown transcript formatting, and AI-generated meeting minutes. The core result is a segments array with speaker_label, start, end, and text, which is rendered into a readable dialogue transcript.
- Full mode: uses vosk-model-spk-0.4 to extract voice embeddings and clusters them to distinguish speakers, with automatic estimation or manual speaker-count control.
- Lite mode: relies on pauses and silence-based heuristics, suitable for quick previews when no speaker model is available.
- Model discovery: checks VOSK_MODEL_PATH, the local ./models/ directory, user-level directories, and system directories, and allows explicit overrides with --model and --spk-model.

Boundaries

The small Chinese model is about 42MB and works well for general conversations, but it may be weaker on domain terminology or accents; the roughly 1.3GB Chinese large model is a higher-accuracy option. For recordings longer than 30 minutes, processing time may be around 0.3 to 0.5 times the audio duration. In meetings with frequent speaker switches, Full mode usually produces more stable segments than Lite mode. The minutes are generated by the Agent from the transcript, making them useful for summaries, topics, decisions, and action items, but they still deserve manual review.