Video Intelligence Analyzer — AI Transcripts & Analysis

Video Intelligence Analyzer

For researchers, educators and creative strategists, turn video speech into one evidence-linked record per video: original transcript, summary, claims, actions and question answers.

Private cloud candidate under validation. Source parsing, schema validation and failure/billing behavior have offline test coverage. Live source retrieval, AI quality and final release acceptance are being evaluated. This Actor remains private.

What does Video Intelligence Analyzer do?

Turn YouTube, TikTok and Instagram speech into a consistent research record: the original transcript, a source-linked summary, takeaways, chapter suggestions, claims, actions, entities, question answers and suggested content copy. Each factual finding includes exact source excerpts and segment references. Source timestamps appear only when the transcript provides them.

This is speech-content analysis. It does not examine video frames, identify people by voice, fact-check external claims, or measure retention and engagement. A generated content brief is explicitly labeled as a suggestion.

Best fit

• Researchers comparing what speakers actually said across public videos.
• Educators turning captioned lessons into structured notes and follow-up questions.
• Creative strategists examining spoken hooks, audience cues and calls to action.
• AI agents that need source-grounded question answers with machine-readable evidence.

For a transcript-only extraction job, select mode: "transcript". For an existing transcript, supply it directly to skip source fetching. All processing focuses expose the same fields and cost controls.

Source coverage

Source	Transcript route	Practical boundary
YouTube videos and Shorts	Public caption tracks	No audio transcription when YouTube captions are unavailable
TikTok	Native captions, then short AI fallback	AI fallback is limited to 120 seconds; longer native captions may work
Instagram reel or video post	Short AI speech transcription	Up to 120 seconds per video; carousel video clocks remain independent
Supplied transcript	Caller text, SRT/VTT text, or timed segments	No source lookup; timing stays unknown when not supplied

Paid long transcription is disabled. Private, login-gated, deleted, speechless or otherwise unavailable videos may return no transcript or a provider failure. Profiles, playlists and shortened TikTok URLs are not accepted. Expand TikTok share links to the full @creator/video/ID URL first. Availability and model quality still require real cloud validation.

Example workflow

An analyst starts with three public explainer videos and asks, “What next steps does the speaker recommend?” The output keeps each video's original transcript, gives answers only when the speech supports them, and attaches source passages. Instagram carousel videos remain separate records so their timestamps are meaningful. The analyst can review the quoted passages before using the findings in a brief.

Quick start input

{
  "videoUrls": ["https://www.youtube.com/watch?v=bjVIDXPP7Uk"],
  "mode": "analyze",
  "focus": "research",
  "questions": ["What actions does the speaker recommend?"],
  "outputLanguage": "English",
  "maxVideos": 1,
  "maxProviderCostUsd": 0.03
}

This URL is a source example, not a verified live-run receipt. maxVideos counts delivered underlying videos, including individual Instagram carousel members. One run accepts up to 20 combined URLs and supplied transcripts. Duplicate canonical source URLs are processed once.

An offline transcript-only example needs no upstream credentials:

{
  "mode": "transcript",
  "transcripts": [{
    "id": "launch-meeting",
    "title": "Launch meeting",
    "segments": [{
      "text": "We will ship on Friday.",
      "startSeconds": 0,
      "endSeconds": 3
    }]
  }]
}

These are synthetic example words. Use text instead of segments for plain text, SRT or WEBVTT. Each timed segment requires text, startSeconds and endSeconds; set both times to null when unknown. Do not provide contradictory text and segments.

Inputs

Field	Default and bounds	Effect
videoUrls	Optional; up to 20 combined inputs	Full public HTTPS video links
transcripts	Optional; up to 20 combined inputs	{id, text} or {id, segments} with optional title/sourceUrl
mode	analyze or transcript	Includes AI analysis, or extracts speech only
focus	general, research, marketing, education	Emphasizes the relevant spoken-content findings
questions	Up to 5; 300 characters each	Answers with evidence or notInTranscript: true
outputLanguage	English; up to 50 characters	Generated analysis language; original transcript and quotes remain unchanged
transcriptLanguage	Optional language code	Requests a source caption language on supported YouTube/TikTok routes; not translation
maxVideos	10; range 1–20	Stops before delivering additional underlying videos
maxProviderCostUsd	$0.10; range $0.001–$1	Bounds estimated provider spend with persistent reservations

Every transcript is bounded to 60,000 UTF-8 text bytes and 2,000 segments. Analysis additionally requires its complete serialized prompt to fit a 110,000-byte envelope. The Actor does not silently truncate long transcripts. Split oversized supplied transcripts into separate jobs. If the transcript is usable but analysis fails, its result is preserved with status: "analysis_failed" and no result event charge.

What data do you receive?

One dataset item represents one underlying video or supplied transcript. Useful fields include:

• transcript.text, transcript.segments, transcript.timing: complete bounded original speech and available timestamps.
• analysis.summary, takeaways, claims, actionItems, entities: generated findings with exact quotes and source segment references.
• analysis.chapters: chapter titles anchored to real source segments; start times are null when unavailable.
• analysis.creative: supported spoken hook, audience and call to action; absent information stays null.
• analysis.answers: your questions, evidence-backed answers or an explicit unknown.
• analysis.contentBrief: suggested title, description and keywords, marked generatedSuggestion: true.
• analysis.coverage: segment counts and truncated: false for accepted analyses.
• provenance: transcript provider, model, fetch time, source language and timing availability.
• exports: run-storage record keys for the Markdown report and available SRT/VTT subtitles.

Illustrative finding shape, using synthetic speech:

{
  "text": "A Friday launch is planned.",
  "quote": "ship on Friday",
  "evidence": [{"segmentId": 0, "startSeconds": 0, "endSeconds": 3}]
}

Quotes are copied directly from selected original source segments after the referenced segment IDs are validated; quotes are not generated by the model. It does not independently prove every generated interpretation. Claims have externallyVerified: false; review passages before relying on them. AI source timing can be approximate phrase timing rather than exact word timing.

OUTPUT and RUN_SUMMARY record the terminal outcome, counts, diagnostics, named events and estimated provider spending. Outcomes distinguish COMPLETE, PARTIAL, VALID_EMPTY, INVALID_INPUT, UPSTREAM_FAILED and CONFIG_ERROR. A provider outage with no useful data fails honestly. A genuine successful lookup with no available transcript is VALID_EMPTY.

Reports and exports

Use Apify Dataset export for JSON, CSV, Excel or other dataset formats. Nested analysis is best preserved in JSON. Download individual reports from the default key-value store using the keys in each row's exports object. SRT and VTT are available only when every source segment has actual timing; the Actor does not fabricate subtitle timing.

API usage after private deployment

This example becomes executable after the private Actor is created and configured. Replace the placeholder token through your environment; do not paste credentials into input or source control.

curl --fail-with-body \
  --request POST \
  --header "Authorization: Bearer $APIFY_TOKEN" \
  --header 'Content-Type: application/json' \
  --data '{"videoUrls":["https://www.youtube.com/watch?v=bjVIDXPP7Uk"],"maxVideos":1}' \
  'https://api.apify.com/v2/acts/khadinakbar~video-intelligence-analyzer/run-sync-get-dataset-items?timeout=240&maxTotalChargeUsd=0.15'

For larger batches, start an asynchronous run, retain the run ID, and read its default dataset after checking OUTPUT and RUN_SUMMARY. A synchronous timeout does not imply that a run stopped: inspect that same run before submitting a replacement.

Use with AI agents

Suggested prompt:

Analyze these video URLs with a maximum of three delivered videos. Answer “Which specific next steps are recommended?” Return each answer with the source URL, exact excerpt, segment ID and available timestamp. Read the terminal outcome and preserve transcript-only partial results. Treat claims as speaker claims, not verified facts.

The structured inputs and outputs are designed for agent use. Hosted Apify MCP discovery, permissions and actual client execution remain untested for this local candidate. Follow the current Apify MCP documentation after private deployment; no standalone MCP server or connected-app write action is included.

Pricing and spending controls

The proposed model is Pay per event plus user-paid Apify platform usage. Candidate rates are $0.12 per analyzed video, $0.02 per transcript-only video, and $0.00005 automatic startup, subject to measured cloud economics. These are configured for private validation; check the live Actor Pricing tab for the effective record. No tier discounts or customer total-cost claim is established.

The analyzed-video event includes its transcript, questions and brief; there is no additional transcript charge for that result. Transcript-only events apply only when that mode was explicitly selected. Each delivered Instagram carousel video is a distinct billing unit. No result event is charged for empty speech, invalid input, failed extraction or a preserved transcript whose analysis failed. Platform usage and the proposed automatic startup event can still cost money on those paths.

maxProviderCostUsd bounds estimated upstream work using published-rate reservations. It is not an Apify event cap or an invoice guarantee. Apify's Maximum cost per run limits event charges; platform usage is separate. Uncertain requests retain their reservation. No paid automatic retry or paid long-transcription fallback is enabled. Interrupted delivery is reconciled against dataset records and named charges; unresolved effects stop instead of being blindly repeated.

Best results and troubleshooting

Use clear speech, existing public captions for YouTube and short speech-bearing TikTok/Instagram clips. Keep questions specific and compare the quoted passages. Missing source timing remains null. A requested caption language with no matching track can produce no transcript. If a source has no speech/captions, provide your own authorized transcript. For source-size limits, split a supplied transcript. A configuration failure requires fixing owner-managed settings before retrying; an ambiguous checkpoint requires inspecting saved delivery and charge evidence first.

Builder’s note

Instagram carousel videos can each start at time zero. Combining them into a single timeline would attach misleading timestamps to findings, so this implementation keeps separate identities and clocks. The same reasoning applies to untimed text: a real excerpt is more useful than an invented timestamp.

Connect the workflow

For dedicated caption collection, start with the public YouTube Transcript and Subtitle Data Scraper, TikTok Transcript Scraper or Instagram Transcript Scraper. To analyze an existing transcript here, map its speech into transcripts[].text, or map timed cues into transcripts[].segments. Keep original source URLs for provenance. This Actor also fetches transcripts itself when you begin with video URLs.

Local setup and responsible use

Run npm ci, then npm test. Put local test input in storage/key_value_stores/default/INPUT.json and run npm start. Remote transcript fetching uses owner-managed SCRAPECREATORS_API_KEY; analysis also uses OPENAI_API_KEY. Keep these in environment secrets. They are never accepted from ordinary input or included in dataset records, outputs or application logs.

Process public or authorized content in accordance with applicable source/provider terms. Transcripts can contain personal information; control access to run storage and set retention appropriate to your use. This independent Actor is not affiliated with YouTube, TikTok or Instagram. Cloud acceptance, human source/data-use review and the user's publication decision remain separate steps.