Record and Summarise Meetings¶
Goal¶
Record a meeting so you can review what was said, find decisions and action items, and keep the original audio alongside a transcript. Add subtitles or an LLM summary when useful.
Why This Pattern Exists¶
Meeting workflows need both fidelity and downstream usability:
- Audio (
.m4a) preserves the source of truth. - Text transcript supports search and editing.
- Subtitle formats support playback, publishing, and accessibility.
- Optional LLM summary speeds post-meeting review.
OpenScreech long-form recording writes incrementally and avoids buffering entire sessions in memory, which is important for long meetings.
Podcast Alternative: One Interface, Separate Microphones¶
For separate microphones on one audio interface, follow Record a Podcast. That walkthrough covers track mapping, individual audio files, and combined transcripts.
Preset: Meeting Standard¶
| Setting | Where | Value | Why |
|---|---|---|---|
| Output directory | Recording Settings | ~/Documents/OpenScreech Recordings (or team folder) |
Centralized retrieval of sessions. |
| Chunk (s) | Recording Settings → Audio Quality | 120 |
Balanced chunk size for readable transcript segments. |
| Bitrate (kbps) | Recording Settings → Audio Quality | 128 |
Good speech quality with moderate file size. |
| Sample rate (Hz) | Recording Settings → Audio Quality | 48000 |
Compatible high-quality recording baseline. |
| Live transcription | Recording Settings → Live Transcription | On |
Continuous transcript updates while recording. |
| Speaker-labelled transcript | Recording Settings → Transcript Outputs | Off unless needed |
Adds a separate chunk-aware transcript.speakers.txt; requires live transcription. |
| Diariser | Recording Settings → Speaker Diarisation Provider | Choose explicitly | Offline VBx estimates speaker count; Sortformer supports up to 4 stable foreground speakers; LS-EEND supports up to 10 and is more sensitive to quiet/overlapping speech. |
| Word-level timing sidecar | Recording Settings → Transcript Outputs | On |
Writes .words.jsonl for advanced tooling and optional LLM processor input. Requires live transcription. |
| TXT output | Recording Settings → Transcript Outputs | On |
Canonical human-readable transcript. |
| SRT output | Recording Settings → Transcript Outputs | On |
Broad subtitle compatibility. |
| WebVTT output | Recording Settings → Transcript Outputs | On |
Browser/web player compatibility. |
| TTML output | Recording Settings → Transcript Outputs | Off (unless required) |
Enable only if your downstream stack needs it. |
| SMPTE-TT output | Recording Settings → Transcript Outputs | Off (unless required) |
Enable for broadcast/media workflows only. |
| Cue split mode | Recording Settings → Subtitle Cue Segmentation | Sentence |
Better readability than strict chunk boundaries. |
| LLM post-processor | Recording Settings → LLM Post-Processor | A dedicated meeting-summary processor | Keeps the prompt and provider configuration reusable and explicit. |
| Processor Input | Recording Settings → LLM Post-Processor | Original Transcript |
Recommended default for normal summaries; no additional output needs to be enabled. |
| LLM Post-Processor Output (.llm.txt) | Recording Settings → Transcript Outputs | On (optional) |
Saves the selected post-processor's result without changing the original transcript. |
| Save LLM Request Payload (.llm-request.json) | Recording Settings → Transcript Outputs | On for debugging, Off for normal use |
Useful for prompt/provider diagnostics. |
| Open finished recording in Finder | Recording Settings → Storage | On |
Speeds handoff immediately after stop. |
Preset: Compliance/Archive Variant¶
Use when you need rich artifacts and traceability.
| Setting | Value |
|---|---|
| Word-level timings | On |
| TXT/SRT/WebVTT/TTML/SMPTE-TT | All On |
| LLM Post-Processing | Optional (often Off for strict source transcript workflows) |
| LLM Request Payload | On if LLM enabled |
Suggested LLM Processor¶
Create or duplicate a processor in the LLM Post-Processing Library, name it clearly (for example Meeting Summary), and use this system prompt:
Summarize this meeting transcript.
Return sections: Decisions, Action Items, Risks, Open Questions.
Keep names, dates, and technical terms exactly as spoken when present.
Do not invent details.
Setup Sequence¶
- Create or select a meeting-summary processor in the LLM Post-Processing Library.
- Open
Recording Settings...from the menu-barSettingssubmenu, then select it underLLM Post-Processor. - Leave
Processor Inputset toOriginal Transcriptfor the standard summary workflow. Optionally choose a saved transcript representation when the processor specifically needs chunk metadata, speaker labels, or word-level timings. - Use
Edit Processor...for provider/prompt changes andTest LLM Post-Processing...with a realistic sample. - Configure the remaining options and output formats in Recording Settings.
- Verify microphone input and model readiness.
- If speaker diarisation is enabled, choose its model and click Download when the selected model reports
Needs download. - Start recording from menu (
Start Recording...). - Stop recording at end; wait for finalization.
- Open the session folder and validate artifacts.
Validation Checklist¶
recording.m4aexists and is playable.transcript.txtexists and is non-empty.- Optional
transcript.speakers.txtkeeps the configured ASR chunks and labels speaker turns within them. - Required subtitle files are present.
recording.diagnostics.jsonexists for QA/debug.- Optional
.llm.txtand.llm-request.jsonare present when enabled.
Tuning Guidance¶
- If transcript sections feel too coarse: lower chunk size (for example,
90). - If too many small subtitle cues: increase sentence-per-cue or word-per-cue settings.
- Speaker-labelled output follows the same cue splitting inside each speaker turn; chunk and speaker boundaries always take priority.
- If summary quality is weak: edit the selected processor to tighten its prompt, change its model, or increase its timeout.
- Keep
Original Transcriptas the processor input unless the prompt genuinely needs another representation. Word-level timing JSONL is substantially larger than ordinary transcript text and can exceed the context window of a local model.
Quick Tips¶
Stop Recordingends capture but can begin diarisation and LLM finalization. Wait for processing to finish before treating the session folder as complete.- Processor Input is optional configuration: existing and standard meeting-summary workflows use
Original Transcriptautomatically. - The meeting summary processor is shared by reference. Duplicate it before changing the prompt if another profile or file workflow intentionally uses the same entry.
See Also¶
Visual Notes¶
- TODO: add screenshot of Recording Settings with the output format toggles visible.
- TODO: add screenshot of cue segmentation controls.
- TODO: add screenshot of completed meeting session folder.