Powered by glolg
Display Preferences Most Recent Entries Chatterbox Blog Links Site Statistics Category Tags About Me, Myself and Gilbert XML RSS Feed
Sunday, May 06, 2018 - 22:33 SGT
Posted By: Gilbert

Pronounced Disagreement

Mr. Ham: *sipping coffee with feet up*

Some me time, for once. Now, what's new? *flips through newspapers* Ah, the Charismatics have run out of patience and are making their move. Should be some jolly good fun - shots getting fired across the bow in the State's Times forum already. Oxford's smacking Parliament's Fake News back on Thum's credentials vis-à-vis their cocked-up cross-examination practices, and yeah, they deserve it. Same old story, it was mainly about who was gonna betray who first, the end.

*turns page* Xi's praising Marx as the greatest thinker of modern times... has China even been Communist for, like, twenty years? Marx's grave costing US$6 a pop to visit should speak for itself. More SJW cultural appropriation accusations, about cheongsams this time. Let's see, high slit, has the legs and figure for it... verdict: not guilty. Eh, at least we got some extra spicy memes out of the rubbish, so it wasn't a total waste.

*continues flipping* Say, trees can talk to each other? Will wonders never cease. Hmm, TRUMP's getting recognized as a legendary emperor by South Koreans, and vaulting up in the polls surpassing Obama at the same point in their Presidencies despite the totally unfair Fake News Barrage, with savvy Millenials swinging to him as his Dragon Energy bro Kanye West all but doubled his black support overnight? Heh, some things just don't change - nice to have some dependable stability, in this uncertain and rootless world. That magical man!

*gets up and stretches* All's well with the world, then. I wonder what the two dolts are up to?

Mr. Robo: Alright, here goes nothing. fefefefefefe... veveveveve... feel any difference in vibration?

Me: Nope.

Mr. Robo: We're totally screwed.

Mr. Ham: Eh, wassup?

Me: We're trying to brush up on our phonetics, if you haven't noticed, thank you.

Mr. Ham: *flipping through notes* How hard can this be? When I last did this, I just threw the schwa in for every vowel in the IPA transcription, sware bland that that was haw ai spake, and threatened to sue if they discriminated against speech impediments. It's smack in the middle of the vowel chart, any error is bounded! And, frankly, you tend to mumble anyway, it isn't even that much of a stretch.



Weird Al can do it, so can you!
[Background, lyrics]


*one slightly flattened hamster later*

Mr. Robo: Bugger, this is hard work. My throat's dry and I'm feeling queasy, and I'm still not getting the right acoustics. I don't think hamster vocal chords are made out for this, human. Mr. Ham's advice is starting to sound better, methinks.

Me: Et tu, Robo?

Mr. Robo: I mean, the rest of the theory isn't that bad, but get a load of this *flips to random page* - the voiced postalveolar fricative is produced by channeling air flow along a groove in the back of the tongue, with the blade of the tongue behind the alveolar ridge and the front of the tongue against the palate... you hafta be, I dunno, some oral kinesthetic genius to keep track of all that movement! My poor throat!

Mr. Ham: *pops up* Well, you can't be no cunning linguist, without being good with your tongue.

*one additionally-flattened hamster later*

Me: Okay, fine, we get to refer to our notes, so it's not all doom and gloom here. Seriously, I wasn't half bad with pinyin transcription when I was like ten years old - though that was probably simpler given that Mandarin is effectively monosyllabic, character-wise, and you only have to deal with the usual five vowels most of the time, unlike English. Or perhaps it's just that there were fewer distractions then? Ooh, new Reddit post!

Mr. Robo: *glumly* It probably isn't that even hard to model pronunciations with LSTMs and the like, but you'd then need a computer for the processing. Hmm, wait, there's this paper from the Naval Research Lab back in '76 - Automatic Translation of English Text to Phonetics by Means of Letter-to-Sound Rules. Seems like they convert regular text with fixed production rules of the form A[B]C = D, where ABC is text, and D is the corresponding phoneme. Could be worth a shot.

Me: Yeah, what do we have to lose anyway? There's the open-source CMU Pronouncing Dictionary for the required IPA data in ARPABET format, which is all we really need.

*some frenetic coding later*

Okay, let's see: one obvious impediment is that the number of letters generally doesn't correspond to the number of phonemes, with 31884 words in the dictionary matching for length, and 82591 having more letters than phonemes. Interestingly, there are 2031 words with more phonemes than letters. Since the production rule is fixed, I guess we just pro-rate letters and phonemes, when they don't match.

ABASH
ʌbæʃ


As such, we gotta collapse duplicate phonemes in the generated pronunciations, and automatically fail the cases where there are more phonemes than letters, but tough luck. That settled, let's just try the max-expectation mappings from the dictionary data, i.e. to predict what the letter B in ABC sounds like, we simply return what it most often sounds like. I suppose we could weigh that according to empirical occurrences, but later.

Mr. Robo: *furiously typing* Gotcha, human.

Me: So how does it do?

Mr. Robo: Not superbly. Backtesting indicates that the system exactly replicates the given pronunciation from the production rules only about 30.4% of the time, 57.3% if you allow for one phoneme difference, and 71.5% if you allow for two. Of course, the dictionary having specimens like "Abplanalp" and "Zwiefelhofer" might have something to do with that; restricting the backtest to the 1000 most common English words raises performance to 39.2%, 68.6% and 83.5% percent respectively.

Me: Well, I dunno, that's not excellent...

Mr. Ham: I tell you, do the schwa thing.

Me: ...but probably good enough. Let's take a look at the transposition errors. The most common by some distance appears to be ʌ → æ, which to be frank seems within normal variation. In fact, the CMU Dict pronunciation of "and", the third most common word in our reference Google corpus, happens to be "ʌnd", which our system happens to predict as "ænd" - which I believe to be the more canonical IPA transcription. Second place, ʌ → ɪ, is possibly more concerning, but then comes ɑ → ʌ, aɪ → ɪ and ɑ → æ, which seem kinda plausible close-range vowel shifts that we probably can get away with.

Mr. Ham: Schwa! Schwa! Schwa!

Mr. Robo: Ah, well, I generated an easy-to-reference set of character-to-IPA charts based on the system. For each letter in each word in normal left-to-right reading order, simply refer to the appropriate chart, and locate the cell with the left column being the preceding letter - dash for none - and the top row being the following letter. For example, for "a" in "and", we use the Middle A chart, and take the value of the cell with column value "-" and row value "N"... which is indeed æ, as expected. Duplicate phonemes are compressed, as previously mentioned.

Me: Of course, it's probably bad to rely too much on this as a crutch, and obviously it doesn't work for other languages, but it's probably a useful fallback guideline.

Mr. Ham: Uh, wait, if you're allowed access to your own reference materials, then why not just use a dictionary if this is what you're concerned about?

*SPLAT*


...did I say something wrong?




comments (0) - email - share - print - direct link
trackbacks (0) - trackback url


Next: 变天


Related Posts:
Intention Manifestation
Mr. Robo's New Profession
Think Thunk
The Week After
Groundwork
Back to top

Copyright © 2006-2025 GLYS. All Rights Reserved.