Local Whisper Transcription on Mac
Velocaption runs Whisper.cpp directly on your Mac. After a one-time model download, the core transcription workflow runs locally without per-minute API keys.
How It Works
-
Download a model (once): open the model selector and pick a size. Models range from tiny (~75 MB, fastest) to large (~3 GB, most accurate). Quantized variants (q5_1) offer a good speed/accuracy trade-off. The model downloads from the CDN and is cached locally in your app support directory.
-
Pick a language: choose from English, Spanish, French, German, Italian, Portuguese, or auto-detect. Auto-detect works well for single-language videos; pinning the language improves accuracy when you know it upfront.
- Start transcription: click Transcribe. A progress bar updates in real-time as Whisper processes the audio. The percentage reflects actual Whisper progress: 2% means audio extraction is done, 2-99% is timestamp processing, 100% is complete.
-
Review the result: the transcript appears in the sidebar with word-level timestamps and confidence scores. Low-confidence words are underlined (yellow, orange, or red) so you can spot errors immediately.
-
Cancel if needed: hit Cancel at any point to stop transcription and discard partial results.
Key Details
Offline fallback for model list. If you lose internet after downloading models, Velocaption uses a cached manifest so you can still see and select your downloaded models. Only the initial model download requires a connection.
Quantized models are the sweet spot. The q5_1 variants run significantly faster than full-precision models with minimal accuracy loss. Start with ggml-small-q5_1 for a good balance.
Word-level timestamps. Every word gets its own start/end timestamp and confidence score. This powers word editing, split, and merge.
What’s Next
- Edit Words in the Transcript: fix Whisper’s mistakes with confidence-guided editing
- Re-Transcribe Without Losing Edits: run a better model without losing your corrections
- Split and Merge Words: fix timing by splitting or combining words
Keep the edit on your Mac.
Choose a license to record, clean, caption, and export without moving private footage into a browser editor.