SpyMeet

Context

The decisions that matter in engineering and business development are made in meetings, and what survives them is whatever someone managed to write down while also taking part.

Constraint

General transcription tools are built for clean audio in one language. A real technical meeting is none of those things: several speakers, poor room audio, domain vocabulary that no general model has seen, and language switching mid-sentence.

Contribution

I built SpyMeet, a Windows pipeline that captures system and microphone audio separately, enhances it, transcribes it through one of three interchangeable backends, and produces structured minutes with a summary, action items, and the full transcript.

Approach

Enhancement runs before recognition rather than after, with loudness normalization, spectral noise reduction, speech equalization, and compression. A custom domain glossary is supplied so technical terms are recognised rather than guessed. Correction, speaker resolution, quality metrics, and summarization are separate stages, so a weak transcript is visible as a quality figure instead of silently becoming a confident summary.

Outcome

A meeting produces a record that can be filed and searched later, without anyone having to choose between participating and taking notes.

Evidence

The work is held in a private repository containing the capture, enhancement, transcription, and summarization stages.