Hacker News Reader

Comment by lolinder | Hacker News Reader

> An Apple spokeswoman blamed the issue on phonetic overlap between the two words, and said the company was working on a fix.

I don't envy the spokeswoman her job, but this is hilariously off base from a linguistics perspective:

* /tɹʌmp/

* /ˈɹeɪ.sɪst/

One word is two syllables, the other is one. One word has a bunch of sibilance and the other has none. One word has close front vowels (including a diphthong!) and the other has a single open back vowel. The only thing they have in common is the /ɹ/, but even that has a slightly different character when in a cluster with /t/.

There's absolutely no way this was an accident of phonetics.

View on HN

Can audio of someone speaking be reliably turned into something like IPA by a computer though? (Very interested if that exists) My impression is that they're working with fewer elements and a lot more fuzziness about what they are.

I doubt we'll ever get a good writeup of exactly how this actually happens, but I'd really like to.

You're correct, I'm using the IPA as a proxy for the audio data they actually use, but it's a decent proxy.

I'd have linked spectrograms if I could find them, but in general sibilants come out very heavily in the audio data, so the complete absence of sibilance in "Trump" alone would make it extremely unlikely that the words would get mixed up. The fact that the words are substantially different in length would also come out in the audio data, as would all of the other features that I mentioned.

If you were to create spectrograms for each word they would be completely distinct from each other to even a human eye across multiple dimensions. There's no good reason to claim that phonetics are the reason why their advanced statistical model connected the two words.

I guess it depends how much of this is running on the device, and how much is running off the device. If its all running on the device then it might be possible for someone to inspect what exactly is going on. If important parts are running off the device then those parts will be a black box.

That's a programmer's understanding of linguistics. You're highlighting descriptive features which are approximations of averages of how we speak English under perfect recording conditions in a specific dialect. There is way more variance and alternation to how people pronounce words than what you're suggesting. How those sounds get interpreted by STT engines is likewise subject to this same variance. IPA isn't code. We don't perfectly reproduce sounds when we're speaking and certainly not to the letter of an approximated IPA transcription.

You didn't correctly transcribe Trump, btw. It's /ˈtɹəmp/ for American English. But it does exemplify my point: there are absolutely people who pronounce Trump with an unrounded vowel and STT engines have to account for this (and they get it wrong for people who are further from the average, outliers).

There are a lot of people who use a triphthong when saying racist as in [ɹɘeɪ]. With a short recording window, it's totally plausible to have these sounds be misinterpreted.

Reply on news.ycombinator.com