← MRM/Projects/Phantom Scribe
Automation / Audio Working prototype Python · local models

Phantom Scribe

The transcription grunt-work, automated end-to-end.

Freelance transcribers spend hours on repetitive processing before the real review work starts. Phantom Scribe runs the whole pipeline — ingest, transcribe, identify speakers, clean, quality-flag, and format to each platform's exact spec — so a human only reviews what actually needs a human.

scribe — process
$ scribe process interview.mp3 --platform rev transcribed (large-v3) · 2 speakers identified fillers cleaned · Rev timestamps [00:00:14] ! 3 low-confidence segments flagged for review interview.rev.txt · earnings estimate logged

Representative output — not a live instance.

What it does
Full transcription pipeline — normalizes audio, detects GPU/CPU, transcribes with selectable model sizes, and identifies speakers via diarization.
Five platform formatters — Rev, GoTranscript, TranscribeMe, Scribie and generic, each with correct timestamps and inaudible/unclear markers.
Seven output formats — plain text, SRT, VTT, JSON, CSV, DOCX and PDF.
Smart post-processing — filler and stutter removal, punctuation, number spelling, and domain vocab (legal / medical / financial / technical).
Quality control & interfaces — confidence-based flagging for human review; CLI, web dashboard, and a watch-folder auto-mode; earnings tracking.
Proof
0
automated tests
0
freelance platforms
0
output formats
0
CLI commands
Python 3.11faster-whisperPyAnnote diarizationFastAPI · HTMXTyper · Rich
🔒
Architecture and verified results shown. Source and binaries are private — this is a capability showcase, not a distribution.

Drowning in
repetitive processing?

I build automation that eats the boring 80% and hands you only the parts that need judgment. Tell me your workflow.

Start a project →