Transcription

RealtimeKit provides two transcription modes powered by Cloudflare Workers AI:

Mode	Model	Processing time	Use case
Real-time	Deepgram Nova-3	During the meeting	Live captions for attendees
Post-meeting	Whisper Large v3 Turbo	After the meeting ends	Transcript files and webhooks

RealtimeKit processes each participant audio stream separately. This helps identify each speaker in the final transcript.

We recommend upgrading to the Workers Paid plan to avoid Workers AI processing limits on the Free plan. Learn more in Billing and Free plan limits.

Real-time transcription

Real-time transcription streams participant audio to Deepgram Nova-3 on Workers AI and sends transcript events to meeting participants during the meeting.

Turn on real-time transcription

You can turn on real-time transcription for participants by setting permissions.transcription_enabled: true in the participant's preset. This lets you decide which participant audio is transcribed. For example, you can transcribe speaker audio without transcribing audience audio.

To update an existing preset, use the Update a preset API:

curl -X PATCH "https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/realtime/kit/$APP_ID/presets/$PRESET_ID" \
  -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "permissions": {
      "transcription_enabled": true
    }
  }'

To create a preset, refer to the Create a preset API reference.

RealtimeKit transcribes audio only for participants who join with a preset that has permissions.transcription_enabled: true.

During the meeting, RealtimeKit streams transcript updates to the client SDK. To access existing transcripts from meeting.ai.transcripts or listen for new transcript events with meeting.ai.on("transcript", ...), refer to Consume real-time transcripts.

Configure transcription settings

The preset controls whose audio is transcribed. The meeting configuration controls how RealtimeKit transcribes that audio. Use ai_config.transcription to set the spoken language, boost recognition for custom terms, and control profanity filtering for a specific meeting.

curl -X POST "https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/realtime/kit/$APP_ID/meetings" \
  -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Weekly product review",
    "ai_config": {
      "transcription": {
        "language": "en-US",
        "keywords": ["RealtimeKit", "Cloudflare"],
        "profanity_filter": false
      }
    }
  }'

Option	Type	Default	Description
language	string	en-US	Language code for transcription
keywords	string[]	[]	Terms to boost recognition (names, jargon)
profanity_filter	boolean	false	Filter offensive language

Real-time supported languages

Real-time transcription is powered by Deepgram Nova-3 on Workers AI.

Nova-3 on Workers AI supports the following languages for transcription:

Language	Code(s)
English	en, en-US, en-AU, en-GB, en-IN, en-NZ
Spanish	es, es-419
French	fr, fr-CA
German	de, de-CH
Hindi	hi
Russian	ru
Portuguese	pt, pt-BR, pt-PT
Japanese	ja
Italian	it
Dutch	nl

Use multi for automatic multilingual detection across all of the languages listed above.

If no language is specified, the model defaults to en-US. For best accuracy, explicitly set the language code matching your audio.

Consume real-time transcripts

Real-time transcription sends interim and final transcript updates to the client SDK. Use interim updates for live captions, and use final updates for transcript history or saved UI state.

Client SDK

JavaScript

// Get transcript entries already received by the client.
const transcripts = meeting.ai.transcripts;

// Listen for transcript updates during the meeting.
meeting.ai.on("transcript", (transcript) => {
  if (transcript.isPartialTranscript) {
    updateLiveCaption(transcript.peerId, transcript.transcript);
    return;
  }

  appendFinalTranscript(transcript);
});

Transcript payload

{
  "id": "1a2b3c4d-5678-90ab-cdef-1234567890ab",
  "name": "Alice",
  "peerId": "4f5g6h7i-8j9k-0lmn-opqr-1234567890st",
  "userId": "uvwxyz-1234-5678-90ab-cdefghijklmn",
  "customParticipantId": "abc123xyz",
  "transcript": "Hello everyone",
  "isPartialTranscript": false,
  "timestamp": 1716700000000
}

Field	Description
id	Unique transcript entry ID
name	Display name of the participant who spoke
peerId	Peer ID of the participant who spoke. Changes if they rejoin.
userId	Persistent participant ID
customParticipantId	Participant identifier set when the participant was added
transcript	Transcribed text
isPartialTranscript	true for interim updates, false for final updates
timestamp	Unix epoch timestamp in milliseconds

---

Post-meeting transcription

Post-meeting transcription generates a transcript after the meeting ends using Whisper Large v3 Turbo on Workers AI and delivers it through a webhook or REST API. To identify speakers, RealtimeKit processes each participant's audio separately before creating the final transcript.

Turn on post-meeting transcription

You can turn on post-meeting transcription when you create a meeting. Set transcribe_on_end: true to generate a transcript after the meeting ends. To also generate a summary automatically after the transcript is available, set summarize_on_end: true.

curl -X POST "https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/realtime/kit/$APP_ID/meetings" \
  -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Weekly product review",
    "transcribe_on_end": true,
    "summarize_on_end": true,
    "ai_config": {
      "transcription": {
        "language": "en"
      }
    }
  }'

Use ai_config.transcription.language to set the transcript language. For supported values, refer to Post-meeting supported languages. If transcribe_on_end is not set, RealtimeKit does not generate a post-meeting transcript.

Post-meeting supported languages

Post-meeting transcription supports Whisper Large v3 Turbo language codes. Omit ai_config.transcription.language to let Whisper detect the spoken language.

Common language codes include:

Language	Code	Language	Code	Language	Code
English	en	Spanish	es	French	fr
German	de	Hindi	hi	Portuguese	pt
Japanese	ja	Italian	it	Dutch	nl
Russian	ru	Chinese	zh	Cantonese	yue

Additional post-meeting language codes

Language	Code
Afrikaans	af
Albanian	sq
Amharic	am
Arabic	ar
Assamese	as
Azerbaijani	az
Bashkir	ba
Basque	eu
Belarusian	be
Bengali	bn
Bosnian	bs
Breton	br
Bulgarian	bg
Catalan	ca
Croatian	hr
Czech	cs
Danish	da
Estonian	et
Faroese	fo
Finnish	fi
Galician	gl
Georgian	ka
Greek	el
Gujarati	gu
Haitian Creole	ht
Hausa	ha
Hawaiian	haw
Hebrew	he
Hungarian	hu
Icelandic	is
Indonesian	id
Javanese	jw
Kannada	kn
Kazakh	kk
Khmer	km
Korean	ko
Lao	lo
Latin	la
Latvian	lv
Lingala	ln
Lithuanian	lt
Luxembourgish	lb
Macedonian	mk
Malagasy	mg
Malay	ms
Malayalam	ml
Maltese	mt
Maori	mi
Marathi	mr
Mongolian	mn
Myanmar	my
Nepali	ne
Norwegian	no
Norwegian Nynorsk	nn
Occitan	oc
Pashto	ps
Persian	fa
Polish	pl
Punjabi	pa
Romanian	ro
Sanskrit	sa
Serbian	sr
Shona	sn
Sindhi	sd
Sinhala	si
Slovak	sk
Slovenian	sl
Somali	so
Sundanese	su
Swahili	sw
Swedish	sv
Tagalog	tl
Tajik	tg
Tamil	ta
Tatar	tt
Telugu	te
Thai	th
Tibetan	bo
Turkmen	tk
Turkish	tr
Ukrainian	uk
Urdu	ur
Uzbek	uz
Vietnamese	vi
Welsh	cy
Yiddish	yi
Yoruba	yo

Output formats

Post-meeting transcripts are available in multiple formats. Use CSV or JSON for application workflows, and use SRT or VTT when you need subtitle files.

Format	Use case
CSV	Spreadsheets and data analysis
JSON	Programmatic access
SRT	Video subtitle files
VTT	Web video captions (<track> element)

Examples

CSV

"1000","peer-123","user-456","cust-789","Alice","Hello everyone"
"3000","peer-234","user-567","cust-890","Bob","Hi Alice"

CSV rows use the following field order: start time in milliseconds, peer ID, user ID, custom participant ID, participant name, and transcript text.

JSON

[
  {
    "startTime": 1000,
    "endTime": 2500,
    "sentence": "Hello everyone",
    "peerData": {
      "id": "peer-123",
      "userId": "user-456",
      "displayName": "Alice",
      "cpi": "cust-789",
      "joinedAt": "2024-08-07T10:15:29.000Z",
      "leftAt": ""
    }
  }
]

SRT

1
00:00:01,000 --> 00:00:02,500
Alice: Hello everyone

2
00:00:03,000 --> 00:00:04,500
Bob: Hi Alice

Consume post-meeting transcripts

After RealtimeKit finishes processing a post-meeting transcript, you can receive the transcript download URL through a webhook or fetch it with the REST API. Use webhooks for asynchronous backend workflows, and use the REST API when you need to retrieve the transcript for a specific session.

Webhook

Configure the meeting.transcript event in RealtimeKit webhooks:

{
  "event": "meeting.transcript",
  "meeting": {
    "id": "bbb8940e-1b97-402a-97d6-2708b7feca41",
    "title": "Weekly sync",
    "endedAt": "2026-06-03T10:30:00.000Z",
    "createdAt": "2026-06-03T10:00:00.000Z",
    "sessionId": "05e57591-d89e-45c9-ae44-08dc1eaad0e0",
    "startedAt": "2026-06-03T10:00:00.000Z",
    "status": "LIVE",
    "organizedBy": {
      "id": "c94c437b-592a-4a39-b9e2-47ef1451e43b",
      "name": "Example organization"
    }
  },
  "transcriptDownloadUrl": "https://example.com/transcript.csv",
  "transcriptDownloadUrlExpiry": "2026-06-10T10:30:00.000Z"
}

REST API

Refer to Fetch the complete transcript for a session.

curl -X GET "https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/realtime/kit/$APP_ID/sessions/$SESSION_ID/transcript" \
  -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN"

Transcript availability

Transcripts are available for 7 days after the meeting ends. Download or copy transcript files before the URL expiry time returned by the webhook or REST API.

Billing and Free plan limits

RealtimeKit's default transcription records each participant's audio track and processes it with Workers AI. Workers AI usage is billed to your Cloudflare account using audio model pricing, which scales by participant audio minutes, not meeting duration.

On the Workers Free plan, Workers AI includes 10,000 Neurons per day. To use more than 10,000 Neurons per day, upgrade to the Workers Paid plan. Workers Paid includes the same 10,000 daily free Neurons, then bills additional usage at $0.011 per 1,000 Neurons.

You can upgrade to the Workers Paid plan in the Cloudflare dashboard under Manage account.

RealtimeKit transcription uses these Workers AI audio model rates:

Transcription mode	Workers AI model	Neurons per audio minute
Post-meeting	@cf/openai/whisper-large-v3-turbo	46.63
Real-time	@cf/deepgram/nova-3 WebSocket	836.36

Data processing and storage

RealtimeKit transcription is a managed transcription workflow. When transcription is turned on, RealtimeKit processes participant audio with Workers AI and stores transcript outputs in RealtimeKit-managed storage.

For real-time transcription, RealtimeKit streams audio from participants with transcription turned on to Workers AI and sends transcript updates to meeting participants during the meeting.

For post-meeting transcription, RealtimeKit processes each participant's audio separately after the meeting ends, creates the final transcript files, and makes them available through a webhook or REST API.

Storage for sensitive workloads

Post-meeting transcription uses RealtimeKit-managed storage for intermediate participant audio tracks and transcript outputs. If you configured external storage for regular recordings, that storage configuration is not used for transcription processing.

For workloads with stricter storage, retention, or processing requirements, use RealtimeKit recording with custom cloud storage instead of managed transcription. You can then run your own transcription pipeline with your preferred infrastructure, access controls, and retention policy.