Problem

I needed a reliable way to transcribe long local audio files without uploading a whole recording blindly or losing context at chunk boundaries. The tool also had to be safe enough for repeated use: no committed API keys, clear outputs, and recoverable per-part transcripts.

Approach

• Built a Tkinter desktop app for direct local use and a Flask web app for browser-based uploads.
• Used pydub and ffmpeg to normalize audio to mono 16 kHz before transcription.
• Split recordings into configurable windows with overlap so long files remain manageable while preserving sentence context.
• Wrote per-chunk transcript files plus one joined transcript with timestamps for review.
• Added a thread-safe job registry in the Flask app so the UI can poll status, progress, logs, and final transcript text.
• Loaded API keys from .env or the UI, avoided committing secrets, and cleaned temporary files after each job.

Results

• A practical transcription utility for .m4a, .mp3, .wav, and other ffmpeg-supported formats.
• Outputs are easy to audit: *_FULL.txt plus timestamped part files.
• The same core transcription logic is reused by the desktop and web interfaces.

Links

• Source repository