# Changelog of Podcast & Audio Transcriber: MP3, Video to Text (`enisbodlli/podcast-audio-transcriber`) Actor

- **URL**: https://apify.com/enisbodlli/podcast-audio-transcriber/changelog.md
- **Full Actor documentation**: https://apify.com/enisbodlli/podcast-audio-transcriber.md

## Changelog

### 0.1 (2026-10-16)

First build for the Store.

- Transcribes audio and video files, podcast RSS and Atom feeds, Apple Podcasts show and episode
  pages, web pages that embed an audio file or name a feed, show names looked up in the Apple
  Podcasts directory, and uploaded files, with the Whisper tiny, base and small models running inside
  the Actor (transformers.js and ONNX Runtime, no outside service).
- Uses the transcript a feed links (`podcast:transcript`: JSON, WebVTT, SRT, HTML or text) when there
  is one, falling back to the next link and then to Whisper.
- Episode filters: episodes per feed, published after, title contains, and only episodes not
  transcribed by an earlier run in the same account. A seen episode counts toward the maximum, so a
  scheduled run never works back through the archive.
- Output: text, timed segments, SRT and WebVTT, detected language, word count and what each result
  was charged. Failed inputs are stored with the reason and are not charged.
- Charges per audio minute (`audio-minute`, `audio-minute-small` for the small model) or per
  publisher transcript (`publisher-transcript`), when each result is stored. Respects the maximum
  charge, the minutes per file and the minutes per run; a restarted run charges nothing twice.
- The tiny and base models run with a quantised encoder and a full-precision decoder, which was
  faster than a quantised decoder; base took about 6.5 CPU seconds per audio minute on 2026-10-16,
  and about 13 on Apify's servers.
- A long file is cut into a multiple of the workers, each part at most ten minutes, so no worker
  waits while another finishes; at 8 GB three parts on two workers had left one idle for 80 seconds.
- Links and text filters over 2,048 characters are refused before any work.