Live Dictation & Captions

openasr live captures audio from a microphone or system loopback and prints final-per-utterance live captions. For packs that declare streaming-partial support, it also emits incremental partial transcripts.

openasr live

By default this opens the system’s default input device (--source mic) and prints human-readable captions to stdout.

Audio sources

Flag	Effect
`--source mic`	Default input device (the default).
`--source system`	Loopback / system audio capture.
`--device <NAME>`	Select a specific microphone by exact or best-effort name.
`--list-devices`	List available input devices and configs, then exit.
`--input-file <PATH>`	Simulate live streaming from a local file (WAV/MP3/MP4/M4A/WEBM/FLAC/OGG) instead of capturing from a device — useful for testing or demos.

Output format

-f / --format controls what reaches stdout:

text (default) — human-readable line-per-utterance captions.
jsonl — one JSON object per line, suited for piping into another tool.

openasr live --source mic --max-seconds 5 --format jsonl

VAD tuning

Voice-activity detection controls when an utterance opens and closes. All thresholds are optional; the defaults come from the loaded model pack.

Flag	Purpose
`--frame-duration-ms <10\|20\|30>`	Realtime frame duration (default `20`).
`--speech-start-ms <N>`	Required speech duration before VAD starts an utterance.
`--speech-stop-ms <N>`	Required silence duration before VAD closes an utterance.
`--pre-roll-ms <N>`	Audio kept before the VAD speech-start point.
`--max-utterance-ms <N>`	Maximum utterance duration before a forced close.
`--no-speech-timeout-ms <N>`	Initial no-speech timeout.
`--energy-threshold <F>`	Energy threshold for the MVP VAD.

Lowering --speech-stop-ms produces shorter, snappier segments; raising it lets the speaker pause mid-sentence without splitting the utterance.

Streaming partials

Packs that declare streaming-partial support emit incremental transcript snapshots between finalized utterances. Two flags control cadence:

Flag	Default	Purpose
`--partial-interval-ms <N>`	300	Minimum interval between partial emissions.
`--partial-window-ms <N>`	3 000	Sliding-window duration for partial snapshot audio.

These flags have no effect on packs that do not support partials.

Session limits

Flag	Effect
`--max-seconds <N>`	Stop after N seconds.
`--max-utterances <N>`	Stop after N completed utterances.

Both are optional; without either the session runs until you interrupt it.

Session export (`--save`)

Save the finalized transcript when the session ends. The file extension controls the export format:

openasr live --save meeting.srt

Supported extensions: .txt, .json, .md, .srt, .vtt.

Flag	Effect
`--save-join-segments`	Join finalized caption segments into one paragraph.
`--save-suggest-title`	Suggest a conservative title derived from the transcript text.

OBS caption integration

For live-streaming setups, openasr live can continuously update a local text file that an OBS Text Source reads via its “Read from file” mode:

openasr live --obs-text-file captions.txt --obs-max-lines 2

Flag	Default	Effect
`--obs-text-file <PATH>`	—	Path to the text file OBS reads.
`--obs-max-lines <N>`	3	Max finalized/revised lines kept in the file.
`--obs-clear-on-start`	off	Clear the file when the session starts.
`--obs-clear-on-stop`	off	Clear the file when the session stops.

Markdown session notes

Write a Markdown note from the session on stop:

openasr live --markdown-note session.md --markdown-suggest-title

Flag	Effect
`--markdown-note <PATH>`	Write a Markdown live-session note on stop.
`--markdown-append`	Append to an existing file instead of replacing it.
`--markdown-title <STR>`	Override the note title.
`--markdown-suggest-title`	Suggest a title from the transcript text.

Diarization

--diarize labels finalized utterances with anonymous speakers (SPEAKER_00, SPEAKER_01, …). Unlike openasr transcribe, live diarization does not accept a --speakers count hint.

Passing --diarize may install the required WeSpeaker speaker-embedder capability pack — the flag itself serves as consent. See the Diarization page for the full speaker-enrollment story.

Model and download flags

Flag	Env	Effect
`-m` / `--model <ID>`	`OPENASR_MODEL`	Model id from the registry.
`--model-pack <PATH>`	—	Use a local `.oasr` pack file directly.
`-y` / `--yes`	`OPENASR_ASSUME_YES`	Download a missing model without the interactive prompt.
`--offline` (alias `--no-pull`)	`OPENASR_OFFLINE`	Never download; fail if the model is not already installed.

Nothing is auto-downloaded silently. The download path only runs when gated by an interactive confirmation, --yes, or OPENASR_ASSUME_YES; --offline disables it entirely.

Runtime posture

openasr live runs the native backend — the same fail-closed, locally-executed engine as openasr transcribe. There is no cloud fallback and no telemetry. Unsupported configurations error out rather than producing fabricated results.

CLI reference — full command overview including transcribe, serve, and model management.
Server API — the GET /v1/audio/realtime WebSocket endpoint provides a server-side streaming path that shares the same underlying event vocabulary (VAD, partial/final transcript segments, speech-boundary).
Diarization — speaker enrollment, embedding packs, and per-speaker labeling details.
Models — available model packs and quantizations.