halcy 
the noise suppression model I am rolling out has a ~10% failure rate with regard to perfect suppression of dog barks, but a ~50% failure rate with regard to perfect suppression of cat meows. this, i believe, aligns with customer expectations
sleepy duck pond appreciator 🦆✨🏳️🌈@halcy my only expectation is there are little snippets in the transcript for when they come through the noise suppression like:
1️⃣ ok team how do we feel about this plan?
🐈 meow~
2️⃣ agreed
very important business, if the cat has approved it, it’s good to go, the cat approval should go on the record
1️⃣ ok team how do we feel about this plan?
🐈 meow~
2️⃣ agreed
very important business, if the cat has approved it, it’s good to go, the cat approval should go on the record
Nick Appleton@halcy but how does it perform when I’m talking while continuously crinkling a plastic chip packet? Or while I’m taking a teams meeting on a laptop on a train platform while some monster diesel train direct from hell screams past me?
@nickappleton surprisingly, these are comparatively easy scenarios that it generally performs okay at, though they are probably rare. We don't have chips or trains, but we have "vacuum cleaner" (20% less than perfect) and "sirens" (25% less than perfect), which should be a good approximation.
the real trouble is interfering speakers for the most part, humans sound a lot like other humans, annoyingly
@halcy does it help if I sing? :) what’s your latency?
@nickappleton I don't have singing validation set numbers for this model, it was doing fine on emotional data, though, which is similar usually
we introduce 10ms latency, it's a prod RTC model and more is intolerable
@halcy We’re still using 20ms for most of ours. Getting good performance with such low lookahead still blows my mind. Nice one 👍