Audio & Video to RAG: Task presets and other exports

Start with the RAG preset for Audio & Video to RAG, replace the public sample URL with your recording, and review the run budget before starting. An Apify account is required; no speech-provider API key is needed. Each Task runs the same Actor with a different input preset. Google Docs delivery and MCP client setup require your configuration.

Choose a Task

Start with the public samples

Most presets use Google’s public 45.3-second, two-speaker commercial speech sample. It demonstrates the input/output shape; it is not a real interview or podcast quality benchmark. The batch adds the public Brooklyn Bridge speech sample. The YouTube preset uses the short zoo clip.

1. Open a Task, review its input, and save your own copy to keep changes.

2. Keep the sample for your first run or replace urls with public HTTPS recording URLs without embedded credentials. Review any language hint when changing the recording.

3. Review run options. Ordinary single direct-audio demos use a $0.10 maximum charge; speaker-label, YouTube and batch demos use $0.15. These are short-demo ceilings, not price quotes. Adjust the budget and timeout for your own recordings. maxTotalChargeUsd is a run option, not an input field; the batch shares one budget.

4. Start once, retain the run ID, and inspect the output. If a response is lost, check the existing run before creating another paid run.

The basic Actor input is:

{
  "urls": ["https://storage.googleapis.com/cloud-samples-data/speech/commercial_mono.wav"],
  "timestamps": "segment"
}

Add "diarize": true for speakers, "exportRag": true for passages, or change timestamps to "word" for word timings. For an interval, add "startTime": 10, "endTime": 30. The presets keep these choices ready for you. These Tasks request no Gemini text add-ons.

Choose a subtitle layout

The audio-to-subtitles preset uses subtitleProfile: "standard" with segment timestamps. Choose short for a narrower layout, or omit the profile to retain the existing exports. Segment or word timestamps are required.

• standard: wraps at 42 Unicode code points per line; reports more than two lines, cues shorter than 1 second or longer than 7 seconds, and reading speed above 20 code points/second.

• short: wraps at 28 code points; reports more than two lines, cues shorter than 0.5 seconds or longer than 4 seconds, and the same reading-speed limit.

Download transcript-000.subtitle-qa.json beside the SRT/VTT, or inspect subtitle_qa in the saved result. Review warnings before publishing. Profiles preserve the source timing: they do not split long cues by time, remove text, or guarantee two lines. These are product presets, not certification against a subtitle standard.

Find the complete result and downloads

The RAG preset uses the RAG Dataset view; the other presets use Overview. These tables do not display every rich output field. Open the full Dataset JSON to inspect segments, words, transcript_llm or rag_chunks when requested.

For a large result, the Dataset row may have result_record_key instead of inline transcript fields. In that run’s key-value store, retrieve the exact key, such as item-000-result, then read the complete result inside its value wrapper. See the API examples for retrieval requests.

Downloads are in the run’s key-value store:

Numbers follow the zero-based input position. Word detail is in JSON; SRT/VTT provide subtitle cues. Check artifact_errors if a download is unavailable; saved JSON can still be present. Inspect error, stop_processing, billing.stop and partial/unknown billing.status before using output or retrying. A completed run can contain failed items; a budget stop can leave later URLs unprocessed.

Whole-recording speaker labels use S1, S2 and so on. With speaker_scope: "chunk", labels such as C1_S1 are local to each chunk; labels from different chunks do not establish the same person’s identity. Labels are not names; review attribution before publishing.

Deliver a transcript to Google Docs with n8n

Use the n8n setup guide and importable workflow. Select your Apify credential, configure Google OAuth and an editable destination document, and review the workflow’s own run budget. Running the Task alone produces Apify output; your n8n workflow performs the destination write. The provided workflow starts the Actor with configured input; adapt it if you want to run your saved Task instead.

One direct-audio transcription and Google Docs append were checked on 23 October 2026 after manual destination recovery. A later complete YouTube-to-Docs workflow passed on 24 October on build 0.2.5 without manual recovery or a second append; see the acceptance report. Test your own destination before activation.

Use the Actor through MCP

Connect your supported MCP client to Apify’s central service using:

https://mcp.apify.com/?tools=fetch-actor-details,andronixmd/speech-to-text-transcriber

Follow Apify’s MCP instructions and client guides for connection and authentication. Keep tokens in client credentials, never in source URLs or Actor input. These presets do not configure ChatGPT or Claude Code, and do not guarantee a separate MCP tool for each Task. The URL exposes the Actor; inspect its current tools and input schema before starting paid work.

A client-neutral request after connecting:

Inspect andronixmd/speech-to-text-transcriber. Transcribe https://storage.googleapis.com/cloud-samples-data/speech/commercial_mono.wav with segment timestamps. Use a $0.10 maximum run charge if the execution tool supports that run option. If it cannot enforce that cap, stop before starting and use the budgeted Task instead. Retain the run handle, poll that run to completion, and retrieve its full transcript and subtitle downloads. Check item errors and billing before calling it complete.

Use the basic input JSON above as Actor input; the cap belongs in execution options when supported. Retrieve the existing run after a timeout rather than starting it again.

For detailed source, pricing, timing and privacy limits, see the Actor reference.