Skip to content
Build Perch
Browse guides

How to transcribe audio locally (no subscription)

Short answer: Right-click any audio or video file in the Files panel and choose Transcribe: Build Perch runs the local Whisper model on your machine and hands back a timestamped transcript. It needs no account or subscription, and nothing is uploaded. The same local model also powers the mic button for live dictation in chat.

Steps

  1. Right-click an audio or video file in the Files panel (wav, mp3, m4a, mp4, mov, mkv, and most other common audio/video containers are recognized) and choose Transcribe.
  2. First time only: Build Perch downloads the local Whisper speech model (about 141 MB, SHA-256-checked) before it can transcribe anything. After that one-time download, transcription never touches the network again.
  3. Review the file name and, for anything over 30 minutes, a time estimate (roughly 5–9× realtime on CPU), then click Transcribe to start.
  4. When it finishes, read the transcript in the dialog and either Save as text (writes a .txt into your workspace’s BuildPerch/documents folder) or Save as subtitles (.srt) for timed captions, or just select and copy the text yourself.

Notes & gotchas

  • It’s speech-only. Whisper hallucinates or produces sparse [Music]-style output on songs, instrumentals, and soundtracks; the dialog warns you before you commit to a long file.
  • The model transcribes English; other languages come out poorly.
  • Nothing about the audio or the transcript is sent anywhere: decoding and inference both run on your CPU. No GPU is required.
  • Live dictation works the same way: click the mic button in the chat input (or press Ctrl+Shift+M) to talk instead of type; text lands at your cursor with no save dialog, and no separate download beyond the one-time model.
  • The AI can also transcribe a file mid-conversation if you ask it to (the transcribe_media tool): same local model, same timestamped segments, useful when you want it to reason about what was said and when rather than just read the text back to you.