Pull quotes and clip times from a podcast transcript

All three read the same input: an SRT or VTT file, or text that has a time at the start of each line, such as the timestamped TXT export ([05:12] Maria: ...). Without times there is nothing to place chapters or quotes against, and the page says so. A Name: label at the start of a line is read as the speaker.

How the quotes are picked

The transcript is split at full stops, question marks and exclamation marks (the Arabic question mark ؟, the Urdu full stop ۔ and the Devanagari danda । count too) into sentences, each with a start and end time (a sentence inside a long caption gets an estimated time from its length). A sentence is a candidate only if its length is in the range you choose: 8 to 20 words, 10 to 40 or 20 to 60. Each candidate then gets a score:

• up to +3 for length, highest in the middle of the range, where a quote card reads best;
• +1.5 for each signal word (never, always, nobody, everyone, everything, nothing, mistake, secret, truth, because, biggest, best, worst, hardest, easiest, learned, realised (or realized), believe, important, problem, surprised, changed, love, hate, wrong, only, most), up to +3;
• +1 if it contains a digit;
• -2 if it starts with a word that points back (and, but, so, or, because, then, it, that, this, they, he, she, which, also, yeah, yes, no, okay, right, well, um, uh, and the contractions it's and that's), since a quote has to stand without the sentence before it;
• -2 if more than about one word in eight is a filler (um, uh, er, hmm, like, yeah, okay, basically, literally, right, so, well), -1 for a question, -1.5 if it has no full stop.

The best-scoring ones are taken one by one, skipping any that begins within a minute of one already taken, so you get a spread across the episode. After the page lists them it says how many sentences were eligible. These numbers are a rule of thumb we chose, not a measurement of what makes a good quote. The signal words, the filler words and the "points back" words are English, so for another language only the length and the punctuation rules apply, and speaker labels are read only when the name starts with a capital letter (Latin or Cyrillic), not in Arabic or Hindi.

Four outputs

"The biggest mistake was the oven." — Dana (0:41)        quotes with times
> The biggest mistake was the oven.                        Markdown block quotes
>
> Dana, 0:41
start,end,speaker,quote                                    CSV for a spreadsheet
00:00:41.000,00:00:44.500,Dana,The biggest mistake was the oven.
1. 00:00:40.000 --> 00:00:45.500  (5.5 s)  The biggest...   clip boundaries for an editor

(The lines above only show the layouts.) The CSV starts every text cell safely, so a quote that begins with an equals sign or a hyphen is not read as a formula by a spreadsheet. If you choose a speaker, only that speaker's sentences are considered; the speaker list is read from Name: labels in your text.

All three read the same input: an SRT or VTT file, or text that has a time at the start of each line, such as the timestamped TXT export ([05:12] Maria: ...). Without times there is nothing to place chapters or quotes against, and the page says so. A Name: label at the start of a line is read as the speaker.

What it cannot do

It cannot tell a funny line from a dull one, notice that a sentence needs the one before it, avoid a quote that is unkind out of context, or make a graphic. It will sometimes put a flat sentence on the list and leave out the line you would have chosen, which is why the list is longer than you need and ordered by time by default. It also does not cut audio: the clip times are for your own editor.

Using the clip times

Open the audio or video in your editor, move to the start time, cut to the end time, and listen to the join: the extra seconds you choose are there so a word is not clipped. For subtitles for a clip, split subtitles can cut an SRT at times you give it, and SRT to CSV puts every caption in a spreadsheet if you would rather scan the whole transcript yourself. To turn the whole transcript into show notes, use podcast show notes from a transcript.