Zoom AI Services Speech Summarization

Use Speech Summarization when one uploaded audio or video file must produce both a transcript and a summary in a single streaming request.

Current API audit: The production endpoint and multipart request were verified on 2026-10-20. Live calls were accepted but returned 402 QUOTA_EXCEEDED before streaming, so the SSE event sequence below comes from the API reference and official sample rather than an observed end-to-end run.

Routing Guardrail

• One media file needs transcript plus summary → use this skill.
• Only a transcript is needed → use Scribe.
• Transcript text already exists and only a recap, summary, or action items are needed → use Summarizer.
• Live meeting audio/video → use RTMS.
• Interactive microphone conversation with spoken responses → use Voice Agent.
• Typed classification, routing, scoring, or image decisions → use Decisions.

Production Contract

POST https://api.zoom.us/v2/aiservices/speech-summarization/summarize
x-api-key: $ZOOM_AI_SERVICES_API_KEY
Accept: text/event-stream
Content-Type: multipart/form-data

Authentication also accepts Authorization: Bearer $ZOOM_AI_SERVICES_API_KEY. Keep the key server-side. The required key scope is aiservices:write:transcribe-summarize:admin.

The official sample reads the same credential from ZOOM_API_KEY. Use one name in your deployment; ZOOM_AI_SERVICES_API_KEY is this skill's preferred name because it distinguishes the API key from Build-platform JWT pairs.

Multipart fields:

• config: JSON string with transcribe and summary settings.
• file: WAV, MP3, M4A, FLAC, OGG, WebM, MP4, or MOV.

Formal limits are 50 MB and 2 hours. Use 8 kHz or higher audio; 16 kHz or higher is recommended. Mono or stereo is accepted.

Workflow

1. Validate the media type, size, and duration locally before upload.
2. Build the multipart request server-side; never expose the API key in browser code.
3. Parse the SSE stream line-by-line, preserving UTF-8 and multi-line data: fields.
4. Handle transcribe_complete before consuming summary frames.
5. Treat each summary frame as accumulated text; emit only the new suffix to a UI.
6. Finish only after done. Treat error or response_state: "failed" as a failed request.

Response Contract

The stream uses these events:

• transcribe_started
• transcribe_complete
• summary_delta
• done
• error

transcribe_complete.transcript may be a JSON string or an object. Summary text may appear under summary_result, summary_result.summarization.overall_summary, nested metadata, summary, or text. See references/api-contract.md for exact parsing fallbacks.

Failure Handling

• 401: invalid key or missing aiservices:write:transcribe-summarize:admin scope.
• 402 QUOTA_EXCEEDED: observed when the account's AI Services credits are exhausted. The catalog does not list 402; treat it as an entitlement signal, not a schema error. Two different API keys returned it on 2026-10-20, Scribe fast mode returned the same code with Free-trial credit exhausted, and there is no Retry-After, so neither rotating keys nor retrying will clear it.
• 405: wrong HTTP method; this operation is POST only.
• 413: file exceeds 50 MB.
• 415: unsupported media type.
• 429: rate limit; honor retry timing.
• 502: upstream transcription or summary failure.
• 503: summary API is not configured.

Do not treat 402 as a malformed request. The request shape can be correct while the account lacks credits. Check the plan and credit balance in Platform Studio; a new project key under the same account will not clear the 402. Decisions and Voice Agent turns still succeeded on the same key, so a 402 on metered transcription or summarization does not make the key unusable everywhere.

Chaining

• One-shot media to typed routing → speech-summarization + Decisions
• One-shot media to localized output → speech-summarization + Translator
• Transcript-only workflow → Scribe
• Existing transcript workflow → Summarizer
• Live meeting media → RTMS
• Real-time spoken conversation → Voice Agent

References

• API contract and SSE parsing
• Official Node.js sample