Times words from the sound they were heard in (DECISIONS #163), for a speech service that
gives the words but no usable times (Gemini's drift by seconds over 30 s). The voice is where
the sound is loud enough (20 ms frames, gaps under 200 ms filled); words share it by length
(a digit counts as a few letters: "15" is said "fifteen"). A phrase end (punctuation) is
pinned to the nearest real pause when one is within 1.5 s of voice, so sentences start when
the speaker does. Measured against the on-device model's times: 0.14–0.25 s mean error on
three English clips (one over music), where Gemini's own were off by up to 11 s.
Times words from the sound they were heard in (DECISIONS #163), for a speech service that gives the words but no usable times (Gemini's drift by seconds over 30 s). The voice is where the sound is loud enough (20 ms frames, gaps under 200 ms filled); words share it by length (a digit counts as a few letters: "15" is said "fifteen"). A phrase end (punctuation) is pinned to the nearest real pause when one is within 1.5 s of voice, so sentences start when the speaker does. Measured against the on-device model's times: 0.14–0.25 s mean error on three English clips (one over music), where Gemini's own were off by up to 11 s.