Podcast to Highlights

Give it a YouTube podcast (or a direct audio/video URL) and get back a timestamped transcript plus the 5-8 moments most likely to work as short clips: start and end times, a suggested title, a hook score and the reasons behind it. Built for clippers, editors and agencies who cut Shorts, Reels and TikToks from long conversations.

How it works

1. Captions first. If the video has YouTube captions (manual or automatic), the transcript is built from them in seconds and you pay only the per-podcast fee.
2. Speech-to-text when needed. No captions, a direct audio file, or transcribe_speech on: the audio is transcribed with a speech model and charged per minute.
3. Highlight detection. Sentences are grouped into 15-60 s windows and scored for hooks (questions, bold claims, concrete numbers, story markers, pacing). Overlapping windows are removed and the top N are returned in timeline order.
4. Optional LLM rescoring. Add your own OpenAI or Anthropic key to get sharper titles and a second opinion on the score. The key is sent only to that provider and never stored or logged.

Output

One dataset record per highlight:

{
  "rank": 1,
  "title": "80% of your time goes to manual work",
  "start": 812.4,
  "end": 846.9,
  "duration_s": 34.5,
  "score": 88.0,
  "reasons": ["hook at the start", "concrete number", "ideal length"],
  "text": "did you know that eighty percent of your time goes to manual work? ...",
  "source_url": "https://www.youtube.com/watch?v=...",
  "source_title": "A conversation with ...",
  "transcript_source": "youtube-captions"
}

The full transcript (segments and words with timestamps) is saved in the run's key-value store as transcript.json; the OUTPUT record summarizes the run (duration, transcript source, minutes charged, LLM status).

Input

Field	Description	Default
url	YouTube video URL or direct audio/video URL	required
highlights	Number of moments to return (3-15)	6
language	Spoken language code (en, es, ...)	auto
transcribe_speech	Force speech-to-text even if captions exist	false
llm_provider / llm_api_key	Optional rescoring with your own key	—
max_duration_hours	Reject longer sources before processing	4

Pricing

• $0.10 per podcast processed (captions route).
• + $0.02 per minute of audio when speech-to-text is used (a 60-minute episode without captions costs $1.30 in total).
• Nothing is charged if the URL is invalid, the source is too long, or processing fails.

Limits

• YouTube and proxies. YouTube blocks most datacenter IPs ("Sign in to confirm you're not a bot"). For YouTube URLs enable Proxy for YouTube with Apify residential proxies (billed to your Apify account, a few KB per run on the captions route) or pass your own proxy. Direct audio/video URLs need no proxy. If YouTube still blocks the request the run fails and nothing is charged.

• Sources up to 4 hours (configurable up to 8).

• YouTube captions are used as published; automatic captions may contain recognition errors. Speech-to-text uses a compact model tuned for speed; timestamps are accurate to about half a second.

• Only public YouTube videos and publicly reachable files. No downloads of video are made in the captions route.

Pairs well with

Content Rewards Campaign Finder: find paid clipping campaigns, then run this Actor on their reference material.

Questions: shortlane.media@gmail.com