Each sentence the Qwen3 voice speaks is now transcribed with Parakeet, the speech recognition Obeya already uses for voice input, and compared with the script. A sentence the voice stumbled over is spoken again, and words it said differently are listed for the agent writing the demo. Below: real takes of the word "Reauthenticate", which Qwen3 cannot read, with its respelling turned off.
Script: "A Reauthenticate button sits below the error."
Script: "The conversation no longer shows a second Reauthenticate button."
Takes that only get one sound wrong ("reauthentipate") are not spoken again: Parakeet gets a letter or two wrong in clean sentences too ("show" for "showed"). They are named instead, as below.
For these six sentences, under the review table:
heard differently (a word the voice cannot read goes into lib/lexicon.qwen3.<language>.json): "Reauthenticate" heard as "reauthentipate", "reauthentipe", "reauthentipates" (scene 1, 2, 3, 6) "second Reauthenticate" heard as "secondary authenticate" (scene 5)
That is how a word like this one ends up in Qwen3's lexicon of respellings. For an existing German demo, rendered afresh with the respelling on:
heard differently (a word the voice cannot read goes into lib/lexicon.qwen3.<language>.json): "Obeya" heard as "Obea", "Obeia" (scene 1, 4, 5, 6) "prüft" heard as "prüfe" (scene 2) "winget" heard as "Wingit" (scene 4) "den Plänen" heard as "dem Plän" (scene 6) "Commits" heard as "Kommentionen" (scene 7)
Most of these are Parakeet misspelling names; the skill tells the agent to add a respelling only where the voice plausibly said the word wrong, and only once the respelling is heard coming back right. None of this goes onto the demo's report page.
Eleven existing demos, narrated afresh with the hearing on:
| Demo | Language | Sentences | Spoken again | Words listed |
|---|---|---|---|---|
| Setup page | German | 19 | 0 | Obeya, Claude, winget, den Plänen, Commits |
| Owner key | German | 17 | 0 | Obeya, Obeyas, curl-Aufruf |
| Unreachable host | German | 10 | 0 | Claude Code, Obeya |
| New task starts | German | 10 | 0 | – |
| Larger dialog | German | 9 | 0 | – |
| Bundled follow-ups | English | 18 | 0 | queue, planned, Obeya |
| Lookup steps apart | English | 10 | 0 | Obeya, Obeya's |
| Token relogin | English | 10 | 0 | Obeya, Claude |
| One Reauthenticate button | English | 9 | 0 | – |
| Demo page links | English | 8 | 0 | shared demo |
| Pause on hover | English | 6 | 0 | – |
A sentence counts as stumbled when 5 or more letters in a row were said differently or added (in an earlier test of 198 takes, every take that repeated itself or babbled reached that, and no clean take did). Letters only missing count from 12 on: Parakeet sometimes drops a word or two of a clean sentence ("ich sage", "shows"), which Whisper heard in the same takes. Hearing a sentence of about 8 s takes about 0.13 s, against about 3 s to speak it.