How long does an hour of audio take?

Local inference runs at a fraction to a multiple of real time depending on hardware; the tab can run in background while you work.

Can it tell speakers apart?

Not automatically in current local models. Conventions like prefixing lines with speaker names during the edit pass are the pragmatic path.

Workflows 6 min read

Meeting Minutes from Recordings: Local Speech-to-Text and Export

The minutes workflow is transcription plus editing: run the recording through local speech recognition, export text for distribution or timestamped SRT for navigation, then fix the 5–10% that models get wrong — names, numbers, terms.

The flow

Extract or keep audio: transcription runs on mono 16 kHz audio locally.
Transcribe with the language set explicitly; speaker labels are not automatic — mark them in editing.
Export TXT for the minutes body; SRT when you need to jump back to the recording while reviewing.
Edit pass: names, figures, action items — check them against the audio at the timestamped lines.

Privacy posture

Everything runs in the browser: audio, model weights inference, output. For confidential meetings this is the entire point — no recording ever leaves the device. Model download happens once and is cached for subsequent meetings.