How the numbers are made
Where the words come from
AudioSage reads only transcripts that shows publish themselves, through the podcast:transcript tag in their own podcast feed. It doesn't record, transcribe or download audio. Each week it reads every show's feed, takes the 10 most recent full episodes that have a timed transcript, and recounts everything. Transcripts are not stored or republished here; each episode links to the show's own copy.
The numbers
- Talk pace: words in the transcript divided by the minutes between its first and last timestamp. The show figure is the median of its episodes. Music, ads and pauses count as time, so read it as "pace of the episode", not of one person.
- Vocabulary variety: the moving-average type-token ratio over 500-word windows (Covington and McFall, 2010), shown as different words per 100.
- Questions: sentences ending in a question mark, per 100 sentences, as punctuated by the show's transcript.
- Signature phrases: every two- and three-word phrase is counted per show and compared with all the other shows using the log-odds ratio with an informative Dirichlet prior (Monroe, Colaresi and Quinn, 2008). The phrases with the highest z-scores are shown, after dropping phrases made only of common words, numbers, and words typical of ads ("promo code", sponsor names). A phrase must appear at least 6 times.
- Topics: how often words from a fixed list for each topic appear, per 10,000 words, compared with the average show here. The word lists are in the source code.
- Who does the talking: only when the show's transcript names its speakers. Generic labels like "Speaker 1" are ignored because they can mean different people in different episodes.
- Guests: only people the show lists as guests in its own feed (
podcast:person role="guest").
What AudioSage doesn't do
- No labels for bias, beliefs, politics or character. These are counts of words, not judgments of people.
- No full transcripts. Quotes are at most 20 words, attributed, and link to the moment in the episode.
- Machine transcripts have errors, and ads can slip through the filters. Treat every number as approximate and check the moment behind it.
Data
All statistics are in stats.json (CC BY 4.0) and the code is open source.