Most podcasts publish their audio in a plain RSS feed: every episode has an <enclosure> with a public MP3 link, a title, a date and show notes. So a podcast transcript is two steps: read the feed, then run speech recognition on the audio. Open-source Whisper does the second step well on clear speech, and it runs on a CPU.
This post pulls podcast transcripts in Python, from a feed, an Apple Podcasts link or a single file, and builds a daily feed of new episodes.
Disclosure: I'm Aurelix, and I built the Podcast Transcript API & Audio to Text tool. It's a paid tool (prices below).
What each row holds
-
text,word_countandsegments: every segment as{start, end, text}in seconds; withincludeWordTimestampseach segment also lists its words, andincludeSrtadds an SRT subtitle file. -
language(detected per 2-minute chunk, so an English ad break does not flip a Spanish show) andlanguage_probability. - The episode as the feed shows it:
podcast_title,episode_title,published,description,image_url,audio_url,episode_guid. -
duration_seconds,transcribed_secondsandminutes_charged, and astatus:ok,partial(stopped at your minute cap, your maximum charge or the run timeout) or why there is no transcript. Rows without a transcript are not charged.
The newest episodes of a show
pip install apify-client
import os
from apify_client import ApifyClient
client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("aurelix/podcast-transcript-audio-to-text").call(run_input={
"rssFeeds": ["https://feeds.npr.org/510289/podcast.xml"],
"maxEpisodesPerFeed": 2,
"maxMinutesPerEpisode": 60,
})
for row in client.dataset(run["defaultDatasetId"]).iterate_items():
print(row["status"], row.get("episode_title"), row.get("minutes_charged"))
print((row.get("text") or row.get("error"))[:300])
An Apple Podcasts show or episode link works in the same field: it is resolved through Apple's public podcast lookup to the show's own feed. Direct audio or video links (MP3, M4A, WAV, MP4, WEBM) go in audioUrls.
A daily feed of new episodes
run_input = {
"rssFeeds": ["https://feeds.npr.org/510289/podcast.xml"],
"maxEpisodesPerFeed": 3,
"incremental": True,
"stateKey": "planet-money-daily",
}
Save the input as a task in Apify Console and schedule it daily. Each run transcribes only episodes an earlier run did not return, so a quiet day costs only the start fee. Add a Slack incoming-webhook URL in slackWebhookUrl to get the new episodes in a channel.
Timestamps for quotes and citations
for s in row["segments"][:5]:
print(f'{s["start"]:7.1f}-{s["end"]:7.1f} {s["text"]}')
Every segment carries its start and end, so a summary or an AI answer can point at the exact minute of the episode.
Limits
- No speaker labels. One text with timestamps; who said what is not marked.
-
A small, fast model. Whisper
baseis good on clear podcast speech and weaker on accents, crosstalk and rare names. - Public audio only. No logins, paywalls or DRM; YouTube and TikTok links are not downloaded (captions for those come from separate tools).
-
Speed. About 4 to 5 minutes of audio per minute of run time at the default 2 GB; give long episodes a long enough run timeout. A run that hits the timeout saves the episode as
partialand bills only the minutes done.
Cost
$4.50 per 1,000 audio minutes on every plan, cheaper than the leading tools, plus a small start fee per run. Each episode is billed for the seconds actually transcribed, rounded up to the next whole minute; a 42-minute episode is $0.19. Failed links, files without speech and episodes an earlier run already returned are free. Apify's free plan includes $5 of credit a month, about 1,100 audio minutes.
Top comments (0)