Skip to content

Long-Form Recording

What This Feature Does

Ordinary recording captures microphone audio to recording.m4a while optionally generating transcript and subtitle artifacts in chunks. It can diarise the completed recording and write a separate speaker-labelled transcript.

Multitrack recording captures separate channels from one physical interface to individual lossless audio files. Track labels identify each microphone, and deterministic processing creates individual and combined transcripts without diarisation.

Design Rationale

This feature is built for long sessions where reliability and recoverability matter:

  • writes audio directly to disk: compressed M4A for ordinary recording, lossless PCM CAF per multitrack channel
  • transcribes in chunks
  • preserves those ASR chunk boundaries in optional speaker-labelled output
  • avoids storing entire multi-hour sessions in memory
  • writes diagnostics for finalization and quality metadata

Multitrack Recording

Choose Start Multitrack Recording... from the menu bar or Home. On first use, OpenScreech opens Recording Settings so you can configure Multitrack Inputs:

  1. Select one physical audio interface with multiple input channels. Aggregate and virtual devices are excluded.
  2. Use Add Track to build your list, even before connecting an interface. Optionally name tracks Mic 1, Alice, or any other label; new unnamed tracks use Track N. Enable the tracks to record.
  3. Move tracks up or down to set their order. Automatic maps the first track to input 1, the second to input 2, and so on. Use a track’s Input menu to override that mapping. Use Check Input Levels to check the enabled microphones without saving audio.
  4. Choose per-track transcripts, a combined transcript, or both, then start recording.

Use the existing Stop Recording action to finish. Routing and track names are locked while recording. The interface must remain connected; OpenScreech reports interruptions instead of silently substituting another microphone. If restart restoration is enabled, it resumes the previous recording mode using its saved configuration.

All selected channels use one sample timeline. Original audio is stored as lossless 32-bit float PCM recording.caf at the interface's sample rate. The ordinary recording bitrate and sample-rate controls do not alter these originals. Optional quiet-bleed reduction applies gentle expansion only to recognition audio, preserving silence and timing. Its threshold is adjustable; attack, hold, release, and attenuation limits are fixed for speech. Test it on your microphone setup before relying on it for quiet speech.

The existing chunk-transcription switch controls whether recognition is queued during capture or starts after Stop. Audio chunks are queued on disk. Final transcript exports are assembled after Stop; this feature does not display live captions. Settings are captured at the start of the session.

Saved channel names and disconnected interfaces

Your track list, names, enabled choices, order, and input overrides are saved independently of the selected interface. You can add, remove, rename, and reorder tracks without connected hardware. Selecting another interface, unplugging it, or refreshing the device list never rebuilds or deletes tracks. Existing saved names are retained when upgrading.

Automatic mappings follow list order, including disabled tracks. Moving or removing a track updates automatic mappings; explicit input overrides stay with their tracks. Duplicate assignments among enabled, available tracks must be resolved before recording.

Click Refresh to update interface and channel availability. If an interface has fewer inputs than your list, the extra tracks remain visible as Unavailable and are skipped for that session. Connecting an interface with enough inputs makes them available again. At least one enabled, available track is required to start. The interface's channel names appear as input information alongside your own persistent track names.

If the interface stops delivering audio during recording, OpenScreech stops capture and finalises the available audio with an error. It does not switch to another microphone or automatically continue the interrupted session after reconnection. Inspect the session's session.json before using an interrupted recording.

When transcription happens

Chunk transcription setting Recognition After Stop
On Completed chunks are processed during recording. Process the final partial chunk, finish any backlog, and assemble exports.
Off Temporary chunks accumulate on disk during recording. Process those chunks, then assemble exports.

Each chunk is transcribed once. Finalisation merges the recognition results and writes outputs; it does not run a second recognition pass. Both settings currently write continuous original tracks plus temporary audio chunks. The worker reads each closed chunk, converts it to mono 16 kHz samples, optionally applies expansion, and passes the samples to FluidAudio. Chunk offsets place word times on the session timeline. Successfully processed audio chunks are deleted; the original recordings remain.

Multitrack session files

The existing recording output folder contains a session folder with this structure:

Recording-<timestamp>-<identifier>/
  session.json
  transcript.txt
  transcript.speakers.txt
  transcript.srt
  ...
  tracks/
    01/
      recording.caf
      transcript.txt
      transcript.speakers.txt
      ...
    02/
      recording.caf
      ...

The root contains combined exports; numbered track folders contain original audio and enabled individual exports. Stable track IDs keep folder names independent of editable names, list order, and channel overrides. session.json records the interface UID, track-to-channel mapping, recorded track IDs, names, capture format, settings, frame count, and any errors. Temporary .chunks files hold queued audio and intermediate recognition; they are removed after successful finalisation and retained on errors.

Existing TXT, SRT, WebVTT, TTML, SMPTE-TT, and timing-sidecar choices apply to the selected output scopes. transcript.speakers.txt is always included for each enabled scope and uses track labels without running diarisation. Word timing is required internally even when .words.jsonl export is off. Missing recognition timing is reported as an error rather than replaced with invented timing in the combined transcript.

The merger uses fixed pause/punctuation/length rules, chronological ordering, and channel order to resolve equal start times. Simultaneous utterances retain overlap labels in text and original word times in JSONL. Caption intervals are split at utterance activity boundaries, with simultaneous track labels combined in one cue. Matching words on different microphones are retained; no LLM is used to merge or deduplicate them.

The optional LLM post-processor runs on the combined output when enabled. Also run the enabled LLM post-processor on each track adds separate requests for individual outputs. The default Original Transcript input already includes track labels. Keep OpenScreech running until finalisation finishes.

Start/Stop Behavior

The following describes ordinary recording; multitrack behaviour is covered above.

Action Result
Start Recording... Creates session directory and begins capture/transcription chain
Stop Recording... Flushes final chunk, finalizes writers, emits end-of-session metadata

When speaker diarisation is enabled, stop first finalizes recording.m4a, then runs the selected FluidAudio diariser over that completed file and aligns its speaker ranges with the existing chunked ASR words.

Session Directory Artifacts

For ordinary recording, per-session folders (Recording-<timestamp>) can contain:

  • recording.m4a
  • transcript.txt
  • transcript.speakers.txt
  • transcript.words.jsonl
  • transcript.srt
  • transcript.vtt
  • transcript.ttml
  • transcript.smpte-tt.ttml
  • transcript.llm.txt
  • transcript.llm-request.json
  • recording.diagnostics.json

Choosing Output Formats

Start with the output your next tool actually needs:

Need Start With
Read, search, or edit the transcript TXT
Subtitles for common editors and players SRT
Captions for a browser-based player WebVTT
Professional timed-text delivery TTML
Broadcast-specific timed text SMPTE-TT
Custom timing analysis Word timing JSONL
A model-generated summary or rewrite LLM post-processor output

See Output Formats Explained before enabling unfamiliar artifacts. Enabling more formats does not improve recognition quality; it produces more representations of the same transcript.

Settings Reference

Open Recording Settings... from the menu-bar Settings submenu. You can also open the same window from OpenScreech Home or App Settings → Management → Recording Settings....

Setting Range/Options Why It Matters
Chunk (s) 15 to 900 Controls transcript segmentation granularity
Bitrate (kbps) discrete options (for example 48-256) Audio quality vs file size
Sample rate (Hz) supported options (for example 22050-48000) Capture fidelity and compatibility
Live transcription On/Off Incremental transcript writing during recording
Word-level timing sidecar On/Off Optional .words.jsonl transcript output for timing-aware downstream workflows; requires live transcription
Speaker-labelled transcript On/Off Adds speaker labels without replacing the source transcript
Diariser Offline VBx / Sortformer / LS-EEND Chooses the explicit FluidAudio speaker model and exposes whether it is ready or needs download
Subtitle outputs TXT/SRT/VTT/TTML/SMPTE-TT toggles Select only required formats
Cue split mode Chunk/Sentence/Multi-sentence/Words Subtitle readability control
LLM post-processor Any entry in the LLM Post-Processing Library Supplies the provider, endpoint, model, prompt, timeout, credential, and wrapping
Processor input Original transcript or one enabled saved transcript representation Sends exactly one selected representation to the LLM as its user message
LLM Post-Processor Output On/Off Sends the completed meeting transcript to the selected post-processor and writes transcript.llm.txt without changing the original transcript
Save LLM request payload On/Off Writes .llm-request.json for request debugging

LLM Post-Processor Workflow

Meeting Recording stores its own processor selection while using the shared LLM Post-Processing Library:

  1. Choose a processor under Recording Settings → LLM Post-Processor.
  2. Use Edit Processor... to open that exact library entry.
  3. Use Test LLM Post-Processing... to inspect and send a request with the live selected configuration.
  4. Enable .llm.txt output when the result is ready for normal recordings.

Processor Input selects one recording representation for the request. Original Transcript preserves the default behavior and is always available. Transcript text output, the speaker-labelled transcript, and the word-level timing sidecar are available only when their matching recording outputs are enabled. The selected text is wrapped by the processor's optional beginning and end tags, then sent as the request's user message; it is not inserted into the system prompt.

Only one representation is sent. This avoids automatically combining large transcript artifacts, which can exceed the context window of local models. OpenScreech does not truncate an oversized input automatically.

Editing a selected processor also affects any profile or file-transcription workflow intentionally using the same entry. Duplicate the processor first when Meeting Recording needs independent behavior.

Diariser Choices

Diariser Useful For Main Trade-off
Offline VBx (Pyannote Community-1) When you do not know how many people will speak May merge similar voices or split one person into multiple speaker labels
Sortformer Background noise and stable identities for up to 4 speakers May disregard quiet, distant, or background voices
LS-EEND Quiet speech, overlap, and up to 10 speakers More prone to false detections and identity changes

Selecting a diariser does not download it. Use the explicit model Download button; OpenScreech shows FluidAudio's listing, download, preparation, ready, and failure states.

Speaker Transcript Segmentation

transcript.speakers.txt uses the existing ASR transcript rather than a second ASR run. Boundaries are applied in this order:

  1. The recorded ASR chunk is the hard outer boundary.
  2. Speaker changes split content within that chunk.
  3. The selected cue mode (Chunk, Sentence, Multi-sentence, or Words) subdivides each continuous speaker turn.

A speaker turn never crosses a chunk boundary. Overlapping speaker ranges are shown as Speakers 1 + 2 [overlap]; speech with no matching speaker range is shown as Unattributed.

Quality and Diagnostics

recording.diagnostics.json captures writer details such as requested/applied settings and finalize errors. This is useful when requested quality parameters are adjusted internally for compatibility.

Scenario Chunk (s) Formats LLM
Internal notes 120 TXT + SRT Optional
Web publishing 90-120 TXT + VTT Optional
Broadcast/compliance 120 TXT + SRT + TTML + SMPTE-TT Often controlled by policy

Common Pitfalls

Pitfall Symptom Fix
Stopping too early after start Tiny or empty transcript Record long enough for meaningful chunks
Enabling every artifact by default File sprawl Enable only downstream-required outputs
Weak summary quality Poor .llm.txt output Edit the selected processor's prompt, model, or timeout
Unexpected shared behavior Another workflow changes after editing the processor Select or duplicate a separate processor for Meeting Recording

Validation Checklist

  • Audio file exists and plays.
  • Transcript file exists and is non-empty.
  • Required subtitle formats are present.
  • transcript.speakers.txt is present when diarisation and its selected model are ready.
  • Diagnostics file exists for QA.

In-App Help Cues For This Feature

In Recording Settings, use ? help for:

  • output format meanings (SRT, WebVTT, TTML, SMPTE-TT)
  • LLM output options
  • reusable LLM processor selection
  • cue segmentation settings
  • diariser selection, model status, and explicit download

Quick Tips

  • Stopping a recording begins finalization. When diarisation or LLM output is enabled, keep OpenScreech running until that post-recording work finishes.
  • Subtitle and timing toggles create additional representations of the same recognition result. Enabling more formats does not improve transcription accuracy.

See Also

Visual Notes

  • TODO: add screenshot of the Recording Settings window.
  • TODO: add screenshot of cue segmentation controls.
  • TODO: add GIF from recording start to saved session folder.
  • TODO: add screenshot of sample session folder contents.