Metaphrasis: Read Audio You Cannot Hear

Two jobs in one skill. Triage a music library without playing it: scan a folder and get one line per track, read a single track for its loudness strip chart and its clean cut points, shortlist for a role, or check which tracks splice together without a seam. Then cut and sew them into one piece.

The second job is QA on narrated video. It returns the transcript, words per minute, rushed and dragging lines, dead air, and a verdict on the mix: clear at 8 dB or more of separation, tight between 4 and 8, buried under 4.

One design rule is worth the price by itself. The mix number never depends on the transcript, it is measured from speech-band energy in the audio. An earlier version read the bed from gaps between transcript segments, and improving the transcript closed those gaps and flipped a perfectly good video from clear to buried.

Requires ffmpeg and ffprobe on PATH, Python with librosa, numpy, scipy and soundfile, and .NET to build the bundled offline Whisper host. Bring your own ggml model.

Where this came from

Deliberon is a Windows desktop app that runs a council of AI agents on a hard decision and hands back a Decision Record: dissent preserved, every claim marked as proved, asserted, or estimated.

Metaphrasis is how its media gets checked, because an AI with no ears will otherwise guess at music in confident prose.

Pay once, it's yours, no subscription. Thirty days of the full council with no account and no card. Runs on your machine with your own model keys.

https://deliberon.com

The rest of my work, organised by material: https://ryan-heltemes.com