Comment by saurik | Hacker News Reader

Comment by saurik | Hacker News Reader

Parent Full thread

The IPA itself isn't really sufficient to encode the sound of a word: you also need to know what language is being spoken; there is a little book you can get of vowel charts for all of the major languages in the world that tries to document the exact vocal position of each symbol. The issue is that there's a continuum of sound that can be generated by the human vocal system, and while humans don't want to differentiate between subtle variations while trying to talk to each other (and so you wouldn't expect two words with massively different meanings based on some extremely subtle and difficult to hear difference), they can notice the difference (which is part of what causes people to have "accents" based on their native languages: because they are pronouncing vowels in some slightly different part of their mouth). IPA only represents the major possibilities, not the specific variants (though there are extended and increasingly complex symbols to try to provide some more accuracy, it is again a continuum and not something you can represent with a small and finite set of symbols per vowel: you'd have to start putting a couple numbers down to represent the sound ;P).

Sadly, I'm having a very difficult time finding a good reference for this online. I know all of this because I spent years studying graduate Linguistics at UCSB while I was trying to get a PhD in Computer Science, and I carried around that little book of per-language vowel charts for a long time ;P.

http://www.antimoon.com/how/english-vowel-chart.htm

> For example, the average British /æ/ is slightly more open (more like /a/) than the average American /æ/.

That said, I typed all of that under an expectation that it was going to work really well, but in practice this website sounds a lot like Dr. Sbaitso, and so the nuance of pronunciation is totally lost anyway ;P.

One minor nit: it's important to distinguish between phonemic and phonetic transcriptions in IPA. The former are what's usually found in dictionaries and does not account for variance in dialects and speakers. The latter attempts to represent utterances as they are actually produced by speakers.

There's still some nuance that is lost in transcription but phonetic IPA transcriptions can achieve a pretty close approximation to the real utterances.

This is true, but you make it sound like this is a response to and even a solution for my comment, when I was actually assuming a phonetic transcription: such transcriptions also vary in how narrowly they define the sound involved, and it is the nuance of sound to which I was commenting, as even in a very narrow phonetic transcription (with tons of marks to try to adjust the sounds of the phonemes), you just can't represent what a native speaker sounds like using these symbols without adjusting for language; and I guess we just have to disagree with how "pretty close" the result is, as when I thought about what I would want to use an IPA->speech tool to accomplish, they all involved vowel charts ;P.

> This is true, but you make it sound like this is a response to and even a solution for my comment

I think it's a suitable response. There is enough allophonic variation among speakers of a single language that undoubtedly someone perfectly reciting a narrow IPA transcription could be taken as a plausible native speaker. Even for your purposes (though I'm still not clear on what type of task your envisioning) the IPA could still be useful as an intermediary layer of abstraction, as in storing a mapping by language of IPA vowel symbols to the exact formants required.

Thanks for taking the time to critique this - it's only something I put together in a few hours for fun. I'm sure somebody with more skill than me would be able to make something like this for multiple languages / accents.

This could actually be quite useful for generating stimuli for artificial language learning experiments (used in psychology and linguistics) where you don't want to model longer-distance effects among speech sounds (avoiding correlated queues), and need a 1:1 correspondence between symbols and sounds. Do you have a command line version?

One other thought: you could compare the output with espeak (http://espeak.sourceforge.net/) or use espeak to generate IPA transcriptions for various languages.

This is really cool! Thanks for sharing with us!

Is this using parametric synthesis? (It doesn't sound concatenative.) Do you have a background in speech, signal processing, or audio, and is this just a passing interest, or something you want to continue to explore?

I've been teaching myself speech algorithms and methods off and on for the past six months ago. Recently I developed a concatenative Donald Trump text to speech engine (I've posted about it in the past), but the samples aren't great and it doesn't use proper unit selection. I'm trying to apply ML to generate a massive set of smooth n-phones that concatenate well together.

I'd definitely like to exchange contact info if you're into speech synthesis long term. My info is in my profile.

In any case, really cool project! :)

It may not be sufficient but it is surely very useful when you have no idea how to pronounce the word. Besides, I think it works quite well for English. When I started learning it I had to rely on the IPA transcriptions every time I looked up new words in a dictionary. Even now, when online dictionaries have nice recordings of the vocabulary spoken by native speakers, I still use it for proper nouns(names etc.) that are not included in dictionaries when I can't find the recorded pronunciation e.g. on Youtube.

It only works for English. Any other language seems to get mangled into English-like syllables.

Are you talking about this project or the IPA generally?

This "project" specifically. It conflates ɾ and t and it ignores most non-english IPA letters.

>> "there is a little book you can get of vowel charts for all of the major languages in the world that tries to document the exact vocal position of each symbol."

What is the name of the book? Do you have a link to it on Amazon or the ISBN for the version you're recommending?

I am pretty sure the book I had was the "Handbook of the International Phonetic Association: A Guide to the Use of the International Phonetic Alphabet".

https://www.amazon.com/dp/0521637511

For a comprehensive site proposing a more accurate representation than IPA, named canIPA, please see http://venus.unive.it/canipa/dokuwiki/doku.php?id=en:start.

Quite true. I guess you would need different "profiles" corresponding to various languages or major dialects so that each symbol is interpreted correctly.

Reply on news.ycombinator.com