Hacker News Reader

Comment by tbabb | Hacker News Reader

It's hard to tell if the speech recognition is working, or even if the microphone is on. If I hum the star spangled banner, it rates me as 79% British.

View on HN

I said the sentence in Spanish and it gave me 62% British. I tried Portuguese, and it gave me 55% British.

Reminds me of the analogy games described by Hofstader in one of his books.

Like "New Zealand is the Scotland of Australia". If i remember correctly, the idea was to take two pairs of countries and construct an analogous mapping between them.

So Spain::Portugal -> UK::USA in this case.

There should be a third mapping there---"neither". I got 99% British with my normal (more or less foreign-accent-free) speach and a whooping 88% British with the heaviest Polish accent I could produce.

The NN (I guess) checks the accent, so it doesn't care about the content of the words. One could speak Spanish in a more or less British accent.

Yes, it would break your words into phonetics, and analyzes whether it's "american" or "british" phonetics which occur more frequently in the dialog. Unless well practiced, your phonetics likely cary over into other languages that you speak.

If it's machine learning-based it probably just doesn't know how to handle weird invalid input like that.`

“Weird invalid input” is an interesting description. That’s the reason I don’t use the voice features on my phone. I have to spend too much time tailoring my input for its expectations rather than Speaking as I normally do, and having it know how to interpret better.

Getting any use out of voice recognition personal assistants is a learned skill (just like effective googling or using the command line) but pretty handy once you get it figured out.

I'd be interested to see a voice-controlled personal assistant with a specific machine-friendly syntax you have to learn instead of it trying to deal with natural language. In exchange for me learning the syntax I'd probably have a much better user experience than fumbling through variations of English.

I find the set of commands I want is usually limited enough that it's not a big deal to learn the expected phrasing.

Well however you want to phrase it it's a common problem. I remember reading about an AI that played Super Smash Brothers and it was really good, but it would start to go haywire if you just ran to the edge of the stage and stood there.

Reply on news.ycombinator.com