# Changelog of Audio & Video Transcriber — Speech to Text API (`andronixmd/speech-to-text-transcriber`) Actor

- **URL**: https://apify.com/andronixmd/speech-to-text-transcriber/changelog.md
- **Full Actor documentation**: https://apify.com/andronixmd/speech-to-text-transcriber.md

## Changelog

### Unreleased — no download fee for an item the cost guard refuses (the four transcript Actors)

- After the audio probe, an item is checked against the operator cost guard before `platform-video-fetch` is charged. When no speech engine fits half of the item's minute price, the item fails unbilled, and the error names each engine's cost, starting with the first one refused. Such an item no longer stops the batch. An engine set that cannot be priced at all, or an unpriced minute event, still stops it.

### Unreleased — AI-agent access docs

- All seven Actor READMEs have the same "Use it from an AI agent (MCP)" section: `search-actors`, `call-actor` with `callOptions.maxTotalChargeUsd`, results through `get-dataset-items`.
- Agent payment without an Apify account is documented: a prepaid Apify AGI token (x402 or MPP, from $1, 14 days). The docs now say that the pay-per-event and no-Standby rules apply only to per-request payments.

### Unreleased — YouTube scraper pricing

- `video-scraped` (YouTube Video Scraper) now uses per-tier prices: FREE $0.005, BRONZE $0.005, SILVER $0.0045, GOLD $0.004, PLATINUM $0.004, DIAMOND $0.004. Run setup (`actor-start`) is unchanged at $0.002.
- The default run memory of the YouTube Video Scraper is 1024 MB, based on the A/B in docs/research/2026-10-31-actor-unit-economics.md.
- The prices take effect on the date Apify shows on the Pricing tab.

### Unreleased — readable output (the four transcript Actors)

- Transcript rows have `paragraphs` (`start_s`, `end_s`, `speaker`, `text`), and each item saves `transcript-NNN.txt` and `transcript-NNN.md` downloads (plus `.translated.txt`/`.md` when translated segments exist). Paragraphs break at speaker changes, pauses of 2 s or more, and length limits; no text is dropped. Free: no price or billing event changed.

### 2026-10-31 — inputs, billing fixes and richer rows (all seven Actors)

- Start from collections: podcast RSS/Atom feeds and Apple Podcasts shows (`feedUrls`, transcriber), YouTube channels and playlists (`channelUrls`, both YouTube Actors) and TikTok profiles (`profileUrls`, both TikTok Actors), with `maxItemsPerSource` and `publishedAfter`. Listing is not charged; listed items are billed like submitted URLs.
- Chain from another Actor: `datasetId` reads a dataset (for example a TikTok or Instagram scraper run) with read-only access; rows without a usable URL become unbilled error rows.
- Scheduled runs: `skipPreviouslyProcessed` with an optional `memoryKey` skips items a previous run already delivered and billed. If the memory cannot be read, the run stops before any charge.
- Billing fixes: incomplete YouTube data (removed, private or bot-checked videos) is no longer billed; the platform download fee is charged only after the file is probed as audio and the run budget covers the fee plus the minutes; recordings without speech return `no_speech: true` and are not charged minutes, up to 30 minutes per run; scrapers charge each row in input order as it is published.
- YouTube: the caption track is chosen without a `language` hint (spoken language, then English, then a sole creator track), transient caption errors are retried once, cue times are clamped to the video length, and failed AI fallbacks name the caption reason.
- Richer rows: video metadata on YouTube transcript rows, creator chapters (`video_chapters`), available caption languages, channel follower counts, video size and more scraper fields; Google Drive and Dropbox share links on the transcriber; clearer errors for image posts, live streams and invalid input.
- Scrapers fetch up to four videos at once, so large batches finish sooner.
- Subtitles: the `short` profile now splits cues to its own limits; `timestamps: none` returns text only.
- Text add-ons can run on Claude when the operator configures `ANTHROPIC_API_KEY` (not enabled on the hosted Actor yet).

### 2026-10-23 — listing and onboarding

- Replaced the Store README with a buyer quickstart, an abridged observed speech result, download links, three recipes and source/capability limits.
- Made hosted Gemini text-add-on unavailability explicit and documented the shared run budget, source classification, failure billing and startup-inclusive event examples.
- Added public API, central MCP and importable n8n onboarding links, with live verification limits stated.
- Moved detailed input, add-on, engine, codec, direct Standby authentication/recovery, signing, metrics and local simulation guidance to the advanced reference.
- Grouped the published input form, aligned supported validation bounds and added output destinations. Existing input names, processing features and event rates are unchanged.