Meeting transcription

Meeting transcription records a meeting and turns it into a timestamped transcript. Speedwave captures audio on the Desktop app host and transcribes it locally with whisper.cpp, so no audio leaves your machine. The container where Claude runs has no audio access.

Beta

This feature is beta, hidden behind a beta features toggle. Treat Polish transcription as a draft rather than a finished record.

What it captures

Each time you record, choose one of three sources: the whole meeting (system audio mixed with your microphone), system-wide audio only, or microphone only, mixed into a single 16 kHz mono stream. The transcript has timestamps but no speaker labels; to learn who said what, send it to Claude and ask it to infer turns.

Turning it on

1. Open Settings -> Beta features and enable beta features to reveal the Meeting transcription tab.
2. Open Settings -> Meeting transcription and download a speech model, Whisper large-v3 or large-v3-turbo, picked for your hardware: roughly 1.5 to 2.9 GiB, SHA-256 verified, network needed once, then transcription runs offline. Downloads cap at 12 GiB total, and only redirects to an allowlist of hosts (Hugging Face, GitHub, and their CDNs) are followed.
3. Open the Meeting transcription tab, pick a language (Polish or English; there is no auto-detect), and press Start to record.

Transcription runs in two passes: a live pass while you record, then a more accurate offline pass that re-transcribes the full recording after Stop and replaces the live text. Controls show which acceleration is in use, Metal or CPU only. If the offline pass fails, for example on missing or empty audio, the session shows a failed: <reason> badge, but the live transcript stays in place to read or send to Claude.

Sending a transcript to Claude

The transcript stays on your machine until you choose to send it. Send to Claude uploads the transcript text, after a confirm dialog, to your configured LLM provider so Claude can summarize it or pull out action items. Only the text is sent, never audio.

Where recordings live

Each recording lives under ~/.speedwave/transcripts/<id>/, as audio.wav and transcript.json kept together, with owner-only file permissions, not encrypted at rest, with no auto-cleanup or retention limit. Long recordings grow the in-memory audio buffer at roughly 115 MB per hour (16 kHz mono); every 30 minutes, Speedwave logs a warning about this to the app logs.

Meeting transcription needs macOS 14.4 or later; Windows captures system audio through WASAPI. On macOS, the first recording triggers a microphone permission prompt, resettable with tccutil reset Microphone pl.speedwave.desktop.audio-capture.

For how this fits the rest of Speedwave’s data handling, see Data and privacy.