Skip to content
Build Perch
Browse documentation

Local transcription & dictation

Short answer: Click the mic button to dictate into any chat, or right-click an audio/video file in the Files panel to transcribe it; both run on a local Whisper model, entirely on your machine, with no audio ever sent to a provider. It needs a CPU with AVX2, standard on essentially any PC built since 2013.

What is local transcription and dictation?

Build Perch ships a local speech-to-text model (Whisper base.en) that powers two separate features:

  • Voice dictation: the mic button in the chat input toolbar. Click it, speak, click again (or use the keyboard shortcut) to stop; the transcript is inserted at your cursor in the message box, ready to send or edit. Press Escape while recording to discard the take instead of transcribing it.
  • File transcription: right-click any audio or video file in the Files panel and choose Transcribe. The app decodes and transcribes the file locally through its bundled media engine, so it isn’t limited to common playback formats: it covers common audio containers plus video containers like MP4, MOV, MKV, and AVI (the audio track is pulled out; the picture is ignored). When it finishes, you choose to save the transcript as a text file or send it straight into the chat.

Both paths use the same on-device model and never send audio anywhere: dictation and transcription work with the app fully offline, on Ollama models and on cloud-connected chats alike. Inside chat, the AI can also transcribe a workspace file itself with the transcribe_media tool (timestamped segments, English) as part of local media editing; that’s the same model doing the same job, called by the AI instead of by you.

How do I dictate or transcribe a file?

  1. Dictate into a chat: click the mic button in the chat input toolbar (or press Ctrl+Shift+M). Speak, then click the button again (its icon changes to a stop square) to stop and transcribe. The text lands at your cursor position in the message box.
  2. Transcribe a file: in the Files panel, right-click a recognized audio or video file and choose Transcribe. Confirm, wait for the progress bar, then pick Save as text file (written into your workspace’s BuildPerch/documents folder, filename editable) or Send to chat (pastes the transcript into the current chat, nothing saved).
  3. First use only: if the speech model isn’t already present, a one-time dialog offers to download it (about 141 MB, verified against a cryptographic checksum before use) before the first transcription; this only comes up if the model wasn’t already installed with the app.

Good to know

  • AVX2 is required. The bundled Whisper build needs a CPU with AVX2 support, standard on essentially any PC built since 2013. A handful of low-end Celeron and Pentium chips lack it; on those, Build Perch shows a clear “not available on this PC” message instead of crashing, and everything else in the app keeps working normally.
  • The speech-to-text model ships inside the installer, so there’s normally nothing to download on first run; the download dialog is a fallback for the rare case where the bundled copy is missing.
  • Transcription is English-only. Other languages, music/singing, and heavily noisy audio transcribe unreliably.
  • Whisper occasionally hallucinates short filler phrases on near-silent audio (e.g. “Thanks for watching!”); Build Perch filters out the exact known patterns before you see the text.
  • No GPU is needed or used; dictation and transcription run as short CPU bursts.