Streaming summaries

Streaming summaries combine file transcription and summarization in a multipart HTTP request. The response contains Server-Sent Events (SSE), including the completed transcript and summary text as it is generated.

Submit a recording#

Use this endpoint:

POST /aiservices/speech-summarization/summarize

Authenticate with your Zoom AI Services credentials. The examples on this page send a token in the Authorization: Bearer header.

This example uploads a project-update recording and selects the corresponding static summary template:

curl --no-buffer --fail-with-body \
  https://api.zoom.us/v2/aiservices/speech-summarization/summarize \
  -H "Authorization: Bearer $TOKEN" \
  -H "Accept: text/event-stream" \
  -F 'config={"transcribe":{"language":"en-US","diarization":false,"profanity_filter":false},"summary":{"language":"en-US","template_id":"project_update","enable_dynamic_template":false,"citations":true,"should_remove_sentiment":false}}' \
  -F 'file=@project-update.wav;type=audio/wav'

Add the config part before the file part.

Let your HTTP client set the multipart Content-Type header and boundary. In cURL, -F does this automatically. --no-buffer displays events as they arrive.

Keep credentials on a trusted backend. Read the POST response as a stream with your HTTP client; a browser's native EventSource cannot submit this multipart POST request.

Configure transcription#

Put recognition settings in config.transcribe. The recognition model is selected automatically and can't be changed.

For accepted media formats and upload limits, see Recording requirements.

Configure summarization#

Put summary settings in config.summary. The summary language can differ from the transcription language.

Automatic template#

Omit template_id and custom_template and set enable_dynamic_template to true, or leave it at its default. The API detects the meeting type and selects a template.

{
    "summary": {
        "language": "en-US",
        "enable_dynamic_template": true,
        "citations": true
    }
}

Static templates#

Set template_id to one of these values. Selecting a static template disables automatic template selection. You can also set enable_dynamic_template to false explicitly.

Custom template#

Set custom_template to a Markdown template. The summary follows the template's headings, structure, and formatting. A custom template overrides automatic template selection. Choose one template mode per request.

{
    "summary": {
        "language": "en-US",
        "custom_template": "## Decisions\n- \n\n## Owners\n| Person | Task |\n| --- | --- |\n",
        "enable_dynamic_template": false,
        "citations": false
    }
}

Handle the event stream#

A successful streaming response uses Content-Type: text/event-stream. Each event has an event: name and JSON in its data: field. A blank line separates events.

HTTP response chunks do not necessarily align with SSE event boundaries. Buffer incomplete events before parsing their JSON.

Transcription events#

event: transcribe_started
data: {"status":"transcribing"}

event: transcribe_complete
data: {"transcript":{"request_id":"3f2b8c1e-7a4d-4e9b-b6f0-1c5d8a2e9b47","duration_sec":5.2,"result":{"text_display":"We agreed to release the documentation on October fifteenth.","segments":[{"id":"utt_0000","start":0.0,"end":5.2,"channel":0,"speaker":"","text":"We agreed to release the documentation on October fifteenth.","words":[]}]}},"latency_ms":5234,"body_bytes":12456,"status_code":200}

transcribe_complete includes the transcription result in transcript, the transcription latency in latency_ms, the size of the transcription result in body_bytes, and the transcription status in status_code. The example shows a shortened transcript; it is not a complete list of response fields.

Summary updates#

Each summary_updated event contains the full summary so far in summary_result. Replace the displayed text with each update. response_state is working during generation and completed in the final update.

Skip empty summary_result values when updating the display. HTML in the summary is escaped: <br> arrives as <br>.

event: summary_updated
data: {"summary_result":"## Decisions\n- Release the documentation on October fifteenth.","citations":null,"response_state":"working","error":null}

event: summary_updated
data: {"summary_result":"## Decisions\n- Release the documentation on October fifteenth.[^1]","response_state":"completed","error":null,"citations":[{"id":1,"marker":"[^1]","start_time":"00:00:00.000","end_time":"00:00:05.200","start_item_index":0,"end_item_index":0}]}

The summary is Markdown. When citations is enabled, the final update adds Markdown footnote markers, such as [^1], and a citations array that links each marker to the transcript. Earlier updates have "citations": null.

Preserve the markers when displaying or storing the summary.

Completion#

event: done
data: {"status":"complete","total_latency_ms":8500,"transcribe_ms":5234,"summary_ms":3266}

The done event reports total processing latency and the transcription and summary stage latencies, in milliseconds. Save the last summary_result; the done payload does not repeat the summary text.

Handle errors#

Check the HTTP status, then read the SSE events. HTTP 200 starts the stream. A successful request ends with a summary_updated event with response_state: "completed", followed by done.

Treat an error event, a summary_updated event with response_state: "error", or a connection that ends before done as a failure. When processing fails, the API sends an error event without done.

An SSE error identifies the failed stage:

event: error
data: {"error":"transcribe failed: connection timeout","stage":"transcribe"}

If transcription fails, the API does not return a partial transcript. If summary generation fails after transcribe_complete, retain the completed transcript and show that summarization did not finish.