Skip to content
Documentation

RTC transcription

RTC can turn a recording’s audio into a transcript after the call. Every participant’s microphone and every shared screen’s audio is transcribed on its own, so each line names the speaker exactly. Two people on one microphone share one speaker label.

Live captions during a call are not available to customers yet. The transcript.segment and captions.* webhooks stay planned.

  • Transcription minutes must be in your plan: Growth and Business include them, and any tier can add packs.
  • Inform participants and obtain the consent your use case needs, for the recording and for the transcript. Your application must show a visible indicator while a recording that will be transcribed is running.
  • Participants who must not be transcribed (for example children, or anyone who declined) join with the transcribe: false claim.

Ask for a transcript when you start the recording, or later for a finished one.

Method and pathScopeResult
POST /v1/recordings with "transcript": truerecordings:writeThe recording starts; its transcript is requested when it ends
POST /v1/recordings/{recordingId}/transcriptrecordings:write202 { "recordingId", "status" }
GET /v1/recordings/{recordingId}/transcriptrecordings:readThe transcript and its status (read)
GET /v1/recordings/{recordingId}/transcript?format=vttrecordings:readThe transcript as a file (formats)
GET /v1/transcriptsrecordings:readYour transcripts, newest first, without their text. Transcripts of recordings in Trash are left out, so a page can hold fewer than limit items; keep following X-Next-Cursor

The request body is optional:

{ "language": "de", "vocabulary": ["CHAMPREP", "Hocuspocus", "Dr. Okonkwo"] }
FieldRules
languageA language code, or zh-Hant for Chinese in Traditional characters. Omit it and RTC detects the language of each participant.
vocabularyHints: up to 100 names, product words or jargon, each up to 64 characters and 1 KB in all. Hints make these words more likely; they do not guarantee spelling.

A transcript goes through these states:

statusMeaning
waitingThe recording is still running. The transcript starts when it ends.
reservingRTC is reserving the transcript’s minutes from your allowance.
queuedMinutes reserved; waiting for a transcriber.
runningBeing transcribed.
doneReady to read.
failedNot made. error and errorCode say why. A failed transcript is free.

A recording has one transcript. Asking again while one exists or is running answers 409 TRANSCRIPT_EXISTS. Ask again for a failed one and RTC runs it again; you may change language and vocabulary. A recording without audio answers 409 NO_AUDIO.

Turnaround is usually well under half the recording’s length plus a few minutes. Use the transcript.completed webhook rather than polling.

GET /v1/recordings/{recordingId}/transcript returns JSON (Cache-Control: no-store):

{
"recordingId": "REC_UUID",
"status": "done",
"language": "en",
"billedMinutes": 42,
"truncated": false,
"durationMs": 2512000,
"model": "faster-whisper-small",
"completedAt": "2026-11-08T12:44:10.000Z",
"segments": [
{ "participantId": "PARTICIPANT_UUID", "identity": "user-123", "source": "microphone", "startMs": 61000, "endMs": 63400, "text": "Let's start with the roadmap." }
],
"createdAt": "2026-11-08T12:00:00.000Z"
}
FieldMeaning
segmentsLines in time order: who (participantId, and your identity when the participant had one), the audio source (microphone or screen_audio), when (startMs, endMs from the start of the recording) and the text
languageThe dominant language, or the one you asked for (zh-Hant for Traditional Chinese)
billedMinutesMinutes charged for this run (billing)
truncatedtrue when the allowance ran out during a run with heavy crosstalk; the text stops there
error, errorCodeWhy a run failed: RTC_ALLOWANCE_EXHAUSTED, BILLING_UNAVAILABLE, CANCELLED, RECORDING_LENGTH_UNKNOWN (the recording’s start or end time is missing, so nothing can be measured or billed), or no code for a processing failure
requestWhat this run was asked for: language (absent when it was detected) and vocabulary, so a retry can send them again

The text is what participants said. Treat it as untrusted user content in your application.

Add format to export a finished transcript as a file. The answer is an attachment named transcript-{recordingId}.{format}. A transcript that is not done answers 409 TRANSCRIPT_NOT_READY for every format except json.

formatContent typeSpeakers
json (default)application/jsonidentity and participantId per segment
vtttext/vttWebVTT voice tags: <v user-123>text
srtapplication/x-subripBefore each cue: user-123: text
txttext/plainOne line per segment: [00:01:01] user-123: text

Where a participant had no identity, the participantId is used.

Transcription understands about 100 languages. They are offered at two levels, from measured accuracy on public meeting and read-speech sets:

LevelLanguages
Generally availableEnglish (en), Spanish (es), Portuguese (pt), German (de), Italian (it), Russian (ru), Japanese (ja), French (fr), Chinese (zh, Simplified characters; zh-Hant for Traditional)
BetaEvery other language, including Arabic (ar) and Hindi (hi)

Beta languages work, with lower and less predictable accuracy. On real meeting audio (several people, spontaneous speech), expect English word error rates around 20 %; read speech and clear single speakers do much better.

GET /v1/capabilities lists the codes under transcription.languages, the generally available ones under transcription.gaLanguages, the formats and the vocabulary limits.

Chinese is written in Simplified characters unless you ask for zh-Hant.

Transcripts use the rtc-transcription-minutes allowance.

  • Unit: one per started minute of the longer of the recording’s length and the total speech across all participants. A normal meeting, where one person talks at a time, is billed for its length. When many people talk at once, the total speech counts instead.
  • Failed runs are free. Nothing is charged for a run that fails or is cancelled; its reserved minutes are released.
  • Reserved first: RTC reserves the recording’s length from your allowance before the transcript is queued. Without enough minutes the request fails free with errorCode: "RTC_ALLOWANCE_EXHAUSTED" and the transcript.failed event lists the packs and upgrades in offers.
  • Crosstalk: while transcribing heavy crosstalk, RTC reserves more minutes as the speech grows. If the allowance runs out, the transcript stops there, is marked truncated: true, and only what was transcribed is billed.
  • If billing cannot be reached, the request waits and then fails free with BILLING_UNAVAILABLE. RTC never transcribes without billing.
  • A finished transcript is always delivered. If billing cannot confirm its minutes afterwards, RTC keeps trying and, after a few hours, records each minute used as usage. Used minutes are never free.

A plan without transcription minutes answers 403 RTC_FEATURE_NOT_IN_PLAN; a used-up allowance answers 402 RTC_ALLOWANCE_EXHAUSTED. Both carry feature: "rtc-transcription-minutes" and the offers that fix it. See plans and billing for the allowances, the pack and pay-as-you-go prices.

A participant whose join token carries "transcribe": false is never transcribed: their tracks are recorded as usual but skipped by transcription. Use it for children, for anyone who declined, or wherever your terms require it. The transcript simply has no lines from them.

Eventdata
transcript.queuedrecordingId, roomId, reservedMinutes
transcript.completedrecordingId, roomId, segments (count), durationMs, billedMinutes, truncated
transcript.failedrecordingId, roomId, error, and for a billing reason code (RTC_ALLOWANCE_EXHAUSTED, BILLING_UNAVAILABLE or CANCELLED) with offers, or RECORDING_LENGTH_UNKNOWN

Webhooks never carry transcript text; read it with the API. See webhooks.

When bring your own storage is on, a finished transcript is delivered with its recording’s files, as transcript.json and transcript.vtt, through the same destinations, path template ({artifact} and {file}), retries and events (recording.delivered with artifact: "transcript:json" or "transcript:vtt"). Recordings in Trash are not delivered.

A transcript follows its recording:

  • Trash: when the recording goes to Trash, its transcript is hidden too (404). Restoring the recording brings the transcript back.
  • Deleted for good: deleting the recording from Trash, or the automatic deletion 30 days after it entered Trash, deletes the transcript with it.
  • Nothing else expires a transcript. It is kept as long as its recording, under your plan’s recording retention.

Copies already delivered to your own storage are yours and are never deleted.

  • CHAMPREP processes recordings and transcripts as your processor, on your instructions.
  • Transcription is self-hosted: it runs in CHAMPREP’s own RTC AWS account in the United States (us-east-1). Audio and text are not sent to any third party, and no sub-processor is involved.
  • Your audio and transcripts are never used to train models.
  • Transcript text never appears in logs, webhooks or CHAMPREP support tools.
  • You are responsible for a lawful basis and the consent of the people recorded and transcribed, for showing a visible recording and transcription indicator, and for laws that protect minors (such as COPPA or FERPA). Use transcribe: false for anyone who must not be transcribed.