Skip to content

v0.20.1

Release date: 2026-11-18

This release fixes cases where stt/verify reported a quotation as disputed although every engine heard the same words, and one case where a quotation with an elision produced a clip covering most of the recording. They were found by checking a podcast case study's quotations against the episode's audio.

Fixed

Spoken forms and word boundaries (stt/verify/)

Engines write the same speech differently. These differences no longer count as disagreements:

  • Informal forms. "gonna", "wanna", "gotta", "kinda", "sorta" and "ok" normalize to "going to", "want to", "got to", "kind of", "sort of" and "okay".
  • A word split in two. Two heard words are joined when together they spell a quoted word, now also when the first is itself a quoted word: "a synchronous" is compared as "asynchronous". They stay apart when both are quoted words on their own.
  • Two words run together. A heard word that spells two quoted words in a row is split: "speedups" is compared as "speed ups". It stays whole when it is itself a quoted word or could split two ways.
  • Broken-off words. A word cut off as the speaker restarted, written with a trailing hyphen ("at ha- at heart"), is dropped unless every token it yields is a quoted word ("run out of- context" keeps "of"; "kinda-" is dropped whole). A broken-off word is never joined to another word or counted as part of a number ("twenty five-" is 20).
  • Locating. Finding a quotation in the timed transcript applies the same rules, so a quoted "speed ups" is found where the transcript has "speedups".

Elisions (stt/verify/)

Words heard across a marked elision ("…") used to cost nothing in the alignment. A quotation with an elision could then match a stretch running on to a later repeat of its last words, minutes away; its clip covered most of the recording and exceeded an engine's upload limit. Alignments now compare by edits first and by words skipped at elisions second: among matches with the same edits, the shortest stretch wins, and a later exact match still beats a nearer match with a misheard word.

Alignment (stt/verify/)

Costs are 64-bit, so aligning a very long reference against an empty hearing returns one deletion per token instead of panicking.

Dependencies

  • Update module dependencies

Documentation

  • The verify guide describes informal forms, word boundaries and elision costs

Tests

  • Regression tests for informal forms, joined, split and broken-off words, and elision spans