Meeting transcription

The meeting transcription feature records audio conversations and produces timestamped transcripts. Speedwave captures audio directly on the host running the Desktop application and transcribes it locally using whisper.cpp, ensuring audio recordings never leave your machine. The container running Claude has no access to audio hardware.

Beta feature

This feature is in beta and requires enabling the beta features switch. Non-English transcripts should be treated as working drafts.

Audio capture and processing

Each recording session supports three audio sources: the full meeting (system audio and microphone input decoded in separate channels), system audio only, or microphone only. An integrated Silero VAD gate suppresses silent intervals, preventing Whisper from hallucinating text during pauses. Generated transcripts include timestamps but do not automatically label individual speakers; you can send the transcript to Claude in chat to infer speaker turns.

Enabling and configuring transcription

1. Open Settings -> Beta features and enable the beta toggle to display the Meeting transcription tab.
2. Under Settings -> Meeting transcription, download a speech model (Whisper large-v3 or large-v3-turbo, matched to your system specifications). The model requires between 1.5 and 2.9 GiB of storage and is verified with SHA-256 checksums. Network access is needed only once during download; subsequent transcription runs completely offline. Total model downloads are capped at 12 GiB, and downloads follow redirects only to trusted hosts (Hugging Face, GitHub, and their CDNs).
3. Open the Meeting transcription tab, select the audio language (Polish or English; automatic language detection is not used), and click Start to begin recording.

Transcription operates in two passes. During recording, a live pass provides an auto-scrolling preview. Once recording stops, an offline pass re-transcribes the entire audio file with higher precision and replaces the preview text. You can also resume finished recordings at any time. The control panel displays the active hardware acceleration method: Metal (GPU) on macOS or CPU on other platforms. If the offline pass encounters an error (such as missing or empty audio), the session displays a failed: <reason> badge while retaining the live draft for review or export.

Sending transcripts to Claude

Transcripts remain on your local machine until you explicitly choose to share them. Clicking “Send to Claude” prompts for confirmation before uploading the transcript text to your configured LLM provider, allowing Claude to summarize discussion points or extract action items. Only textual content is transmitted; audio files are never uploaded.

Local storage of recordings

Each recording is saved under ~/.speedwave/transcripts/<id>/ as a pair of files: audio.wav and transcript.json. These files are accessible only to the owner account, stored unencrypted at rest, and are not subject to automatic deletion. During extended recording sessions, the in-memory audio buffer expands at approximately 115 MB per hour (16 kHz mono), with Speedwave logging a diagnostic reminder every 30 minutes.

Meeting transcription requires macOS 14.4 or later. In Windows, system audio is captured via WASAPI. On macOS, the first recording triggers an operating system microphone permission prompt, which can be reset using tccutil reset Microphone pl.speedwave.desktop.audio-capture.

For details on how transcription data is handled, see Data and privacy.