@vikramp7470 Thank you! Something that is in truth the guts of your query: it isn’t natural WebGPU. When WebGPU is not to be had (previous Intel, base M1, system already loaded), it robotically switches to a lighter backup mode. We in point of fact designed for the ones machines first.
The principle thought on reminiscence: it by no means grows with the period of the assembly. A three-hour name makes use of about the similar reminiscence as a 3-minute one. The audio flows thru a fixed-size “field” (round 1 minute of sound, reserved as soon as at startup). Not anything piles up, so there is no leak that at last crashes the tab. And the transcription runs in a separate lane, by no means at the web page itself, so Meet/Zoom/Groups do not lag whilst it really works.
On no longer crashing the web page: if the system cannot stay up, the sound simply waits in that field as an alternative of stacking up ceaselessly. If it in point of fact overflows, we drop a small little bit of the transcript as an alternative of crashing. Similar with warmth: if the chip slows down, transcription slows down somewhat, that is it, not anything breaks.
Glad to move deeper on any of it!
