search

stts-se / chromedictator

Demo app for testing Google Chrome's ASR API
MIT License
0 stars 0 forks source link

Sometimes the recogniser stops for no apparent reason #52

Open HannaLindgren opened 5 years ago

HannaLindgren commented 5 years ago

Sometimes the recogniser stops for no apparent reason, although auto reload is enabled.

jensedlund commented 5 years ago

For the task of the users of this in our current context, this isn't a low prio bug but rather the opposite. The interpreters are highly likely to miss that the rec has stopped, and lose a few utterances, which is enough to throw them off the track completely once they notice.

I'll try to pinpoint when and why this happens and provide the info.

HannaLindgren commented 5 years ago

Hypothesis: it may have to do with long utterances. When we do internal tests, we typically use pretty short utterances, but maybe the interpreters test with longer utterances?

I had a session yesterday when I got the impression that longer utterances where more problematic. Example below.

image image

jensedlund commented 5 years ago

@HannaLindgren: That's perfectly possible - the interpreters are working under high stress, and also aren't trained for this situation. I've seen them, on several occasions, simply forgetting to either pause or go "punkt" and run on until the system makes something funny - in a previous version this even caused a crash, if I remember correctly.

But this is not the only time it occurs. I've definitely seen it happen after a pause induced send that followed on perfectly sensible-length utterances.

HannaLindgren commented 5 years ago

@jensedlund OK well maybe it doesn't explain everything but if it helps us reproduce the issue that's at least one step forward...