[freddie_mercury]

A quick Google tells me Perplexity doesn't support my language, so no a voice query to Perplexity won't work.

"Perplexity is now available in Korean (한국어), Japanese (日本語), German (Deutsch), French (Français), and Spanish (Español)"

It's missing 3 of the top 5 languages spoken in the world today. It only has 3 of the top 10 most spoken languages in the world.

Also everyone in my country drives scooters and I don't see how voice commands are supposed to work with that kind of setup.

[godelski]

  > Perplexity is now available in Korean (한국어)

Wow, usually Korea takes awhile to get things.

  > Also everyone in my country drives scooters and I don't see how voice commands are supposed to work with that kind of setup.

Even if they were supported I would suspect that you would have some issues. My partner is Korean (I'm not) and when she goes back to Korea and we video chat while she's on the subway then her airpods' "voice isolation" systems actually amplify background noise rather, drowning her out, rather than doing the opposite (its intended usage).

Being an ML researcher... I have some suspicions as to what is the root cause. I can't tell you exactly without having access, but I can tell you that it is VERY common for models to be trained without sufficient augmentation, let alone without sufficient diversity in data. So it is fairly common for the system to not be useful in certain environments. Worse, when you hyper focus on maximizing test data results you are extremely likely to perform data leakage (this is quite common and at this point I'd say you should assume it exists) and not enough evaluation on completely held out sets. Especially data which is significantly dissimilar to what you trained on (even when it is "in distribution"[0])

[0] This term can be a lot of things... so it must be interpreted in context.