How blobbyblo/cqam-transcription works
Three Main Zones
• Zone 1: External Dependencies
• FFmpeg (audio codec/format conversions)
• Hugging Face Hub / Whisper models
• Pyannote speaker‐diarization models
• Local file system for input/output storage
• Zone 2: Processing Pipeline
• Audio Preprocessing (utils/audio_utils.py)
• Speaker Diarization (diarization/diarize.py)
• Transcript Generation (transcription/whisper.py)
• (Optional) Manual Correction module
• Zone 3: UI Layer
• Tkinter GUI (gui/app.py)
• Pygame audio playback component
• Main entrypoint (main.py)
Components & Interactions
• FFmpeg: called by Audio Preprocessing for format conversion
• Whisper models: fetched from HF Hub, loaded by transcription module
• Pyannote models: loaded by diarization module
• Local FS: project root, .env for HF token, input/output folders
• Audio Preprocessor
– Inputs: raw .wav/.mp3 → normalizes, resamples → outputs cleaned audio
• Diarization
– Inputs: cleaned audio → assigns speaker segments → outputs time-stamped speakers
• Transcription
– Inputs: cleaned audio (+ speaker segments) → outputs text with speaker labels
• Correction (if used)
– Inputs: transcript → user edits → outputs final transcript
• Tkinter App
– File browse controls, start/pause buttons, volume slider
– Loads pipeline via subprocess or direct imports
• Pygame Audio Engine
– Handles playback, reports current time to GUI
– Signals UI to highlight transcript segments in sync
Simplification
• Omit test folders, setup scripts, and low-level utility internals.
• Do not show every file; aggregate by function (e.g., “Audio Preprocessing” rather than individual helper files).
• Focus only on how audio/data flows and which external systems are involved.
• The three logical layers you requested
• All major external tools and services
• The sequence of processing steps
• The UI interactions for playback and transcript display