Managers needed transcripts for meeting notes and action items, but confidential discussions could not go to public AI
services and manual transcription took time.
Built a macOS and Windows desktop app with Tkinter/ttk. Apple Silicon uses MLX Whisper; Intel Mac and Windows use Faster Whisper with CTranslate2 on CPU. Users choose a file, language and TXT/SRT output. Background workers keep the interface responsive and show progress and detailed errors. The first run downloads dependencies and the model; later transcription runs locally.
The stack includes Python 3.10+, Hugging Face Hub for model downloads and cache, and FFmpeg for audio/video decoding (installed through Homebrew when needed on macOS). Apple Silicon uses mlx-community/whisper-large-v3-turbo; CPU mode uses Faster Whisper, CTranslate2 and large-v3-turbo. Model download progress is weighted by file size so the percentage does not reset between files. Processing runs in worker threads with a GUI event queue; errors show captured output and a full, copyable Python traceback.
According to the case materials, a one-hour meeting is transcribed in 3–5 minutes instead of 1–2 hours of manual work. Audio is processed locally; the materials also report saving 6–8 hours of team documentation time per week.
| Metric | Result |
|---|---|
| Transcription of a one-hour recording | 1–2 hours manually → 3–5 minutes automatically |
| Audio sent to external services | 0 bytes |
| Team documentation time per week | 6–8 hours → minutes of review |
| Windows without a GPU | Runs on CPU (CTranslate2) |
Speed and time savings are stated in the case materials; actual performance depends on the recording and computer.