Whisper

Transcribe the audio you have selected into text notes on the canvas, on your own machine. Nothing is uploaded: the model is downloaded once and every run after that is local.

What it does

Select one or more audio clips, run Whisper: Transcribe Selection, and each clip's speech comes back as a text note anchored below the clip it came from. The transcript is also posted into the chat that ran the command.

Clips are transcribed as you hear them — the visible slice of the clip, with its fades, gain, pan and its whole effect chain applied. That is the same offline render path as export, so what Whisper hears is what your listener hears, not the raw file underneath.

Every transcript in a run lands as a single edit. Transcribe four clips and one ⌘Z takes all four notes back.

Using it

1. Select the audio clips you want.
2. Run /whisper:transcribe-selection in chat.
3. Pick a language in the chat card, or leave it on Auto detect.
4. Press Process, or Cancel to leave the selection unchanged.

The chat card offers Auto detect plus English, Spanish, French, German, Italian, Portuguese, Russian, Japanese and Chinese. Auto detect is the default and matches how Whisper behaves with no language set.

A selection with no audio clip in it says so in chat. Palette and Extensions window launches direct the user to chat; Whisper has no separate window.

What it needs

• Your project, for writing — it creates the text notes.
• Your audio, for reading — it renders the selected clips to transcribe them.
• One network download. On first use it fetches the base model (141 MiB) from huggingface.co. The download goes through the host, against the host's own allowlist, re-checked on every redirect hop, and its SHA-256 is the one in the signed catalogue — not one the extension asserts about itself at runtime. After that the model is cached and the extension needs no network at all.

The chat job card reports model download, loading, rendering and transcription progress and can cancel the job before it creates notes.

Platforms

macOS, Apple Silicon and Intel.