Mark Liberman has a Log post about a recent paper, Jialu Li and Mark Hasegawa-Johnson’s “A Comparable Phone Set for the TIMIT Dataset Discovered in Clustering of Listen, Attend and Spell” (pdf). The abstract reads:

Listen, Attend and Spell (LAS) maps a sequence of acoustic spectra directly to a sequence of graphemes, with no explicit internal representation of phones. This paper asks whether LAS can be used as a scientific tool, to discover the phone set of a language whose phone set may be controversial or unknown. Phonemes have a precise linguistic definition, but phones may be defined in any manner that is convenient for speech technology: we propose that a practical phone set is one that can be inferred from speech following certain procedures, but that is also highly predictive of the word sequence. We demonstrate that such a phone set can be inferred by clustering the hidden nodes activation vectors of an LAS model during training, thus encouraging the model to learn a hidden representation characterized by acoustically compact clusters that are nevertheless predictive of the word sequence. We further define a metric for the quality of a phone set (the sum of conditional entropy of the graphemes given the phone set and the phones given the acoustics), and demonstrate that according to this metric, the clustered LAS phone set is comparable to the original TIMIT phone set. Specifically, the clustered-LAS phone set is closer to the acoustics; the original TIMIT phone set is closer to the text.

Mark says:

As exemplified above, the TIMIT phonetic transcriptions often reflect expectations from the formal dictionary-based pronunciation standard, which is influenced by the spelling even before any continuous-speech reductions set in — so matching TIMIT’s performance on this paper’s “metric for the quality of a phone set (the sum of conditional entropy of the graphemes given the phone set and the phones given the acoustics)” should not be all that difficult. Still, no one has ever done it before, so this research is an important contribution.

The relationship between phonetic variation and lexically-stable phonological categories remains an open theoretical question, in my opinion, but work like this is one very useful direction of inquiry.

This sounds like it must be important, but it’s been so long since I had anything to do with that kind of linguistics that I’m only vaguely aware of how it works. But it sounds like a useful alternative to the usual transcriptions.


  1. Off topic, though related to Language Log, where I posted on the origin of “on the fritz.” I proposed it transitioned from colloquial friz or fritz meaning frozen to meaning not working. Comments are closed on that old guest post, but I found another relevant text. A young Indiana fellow was sent by the Army Signal Corps to a telegraph outpost in Alaska. On November 8, 1901 he wrote to a friend in Indiana, and a newspaper published it. Crawfordsville [IN] Weekly Journal, Jan. 2, 1902, page 7, col. 3. He described the cold, the ice, and having little to do. He wrote: “This country is on the Fritz proper.”

  2. Stu Clayton says

    He capitalizes the word, as if he thought it was a German name. This attributes it to Fritz of the Katzenjammer Kids. The strip first appeared in 1897 and “became wildly popular”.

  3. The 1902 newspaper does capitalize Fritz, and maybe the 1901 letter writer did also. Both upper and lower case uses appear (some given
    E.g.., possibly related uses of the transition [some may be of debatable relevance], in addition to those in OED June 2014 and HDAS:

    1880 “married or ‘fritz to’ the dark eyed senoritas”

    1886 “a friz nose”

    1891 “Fort’nate they [hands] friz to the oars”

    [1892 “Jimmy the Bunco” schemes to get a Thanksgiving dinner; the lemonade comes with “friz.” “‘I dunno as I cares on the friz,’ murmured ‘the Bunco’ thoughtfully. The word bore too close a resemblance to his general state of being.”]

    1897 “friz up all de creeks”

    1901 “getting t’ be on de Fritz”

    1901 “For everything ‘t was frizable, that year was friz.”

    1902 “Would Santa Claus be on the fritz/ if we never had snow?” [Ironic effect of lack of ice?]

    1904 Life in Sing Sing. “Fritzer. Not good.”

    1905 “He’s on the friz.” [Baseball player slump.]

    1905 “business goes on the fritz.”

    1905 “good manners done friz up”

    1905 four wagons “all to de fritz”

    1906 “is he straight, or is he on de fritz?”

    1908 “Deep breathing is the thing for you if you are on the friz.”

    1908 poem, “friz” rhyming with “wits.”

    1908 Munsey’s. “our fat leading lady was on the friz”

    1909 “show is on de fritz”

    1912 “A poor man is friz out these days. Friz out, I say.”

    1912 “All the religion ‘ll be friz out of this c’mmunity.”

    1912 “I may talk on de fritz” [but won spelling bees]

  4. Surely “friz” as a dialect conjugational form of “freeze” is a quite different phenomenon?

  5. As a dialect conjugational form of “freeze,” “friz” can be seen as a different phenomenon than the formation of the “on the friz/fritz” idiom, if you wish.
    But, are you claiming that the former *cannot* have in a sense led into the latter, and *”Surely”*?

  6. Well, it’s not outside the realm of possibility, I guess.

  7. I think “not outside the realm of possibility” is as far as I’m willing to go, too.

  8. Stu Clayton says

    Then it’s possible, i.e. inside the realm of possibility. Unless you are suggesting as a third possibility (!) “sitting on the boundary between the possible and the impossible”. A more conventional way of expressing that is “I haven’t a clue”.

  9. No, that would suggest no sense of relative probability, whereas “not outside the realm of possibility” implies “very unlikely indeed.”

  10. David Eddyshaw says

    To Plato, I said: “Tell me, P:
    What are these pink rats which I see?”
    With a soupçon of pride,
    He politely replied:
    ‘Τὰ μεταξὺ τοῦ ὄντος καὶ μή.’

  11. Stu Clayton says

    So you think of the boundary between the possible and the impossible as a band of non-zero width. You imagine placing your bets there, closer to the one or the other side of the band – as if betting on rouge or noir at roulette. You’re betting on the probabilities of two possibilities (“possible” and “impossible”). That’s pretty neat, though meta to a fault.

  12. Stu Clayton says

    μεταξὺ is older than I thought !

  13. David Eddyshaw says

    Of course. Indispensable aid to philosophising.

  14. Of course one can reject without offering an alternative. But, speaking of relative probability, does anyone reading care to offer a specific origin proposal–not just “nobody knows”–of “on the fritz” that they consider relatively more probable than what I proposed? (Maybe note that the earliest so far reported uses *may* not appear to refer to machines nor Germans nor cartoons.)

  15. I’m afraid “nobody knows” is the best one can do for a great many words and expressions, especially slang, and it’s a great mistake to think any explanation is better than none. I’m perfectly comfortable with “nobody knows” and feel no impulse to think up something myself. Better to be ignorant than wrong.

  16. It seems to me that while “possible” and “not outside the realm of possibility” are logically equivalent, there is a great deal of difference between them in terms of their scalar implicatures.

  17. Need I say?–I do not think any explanation is better than none.
    I merely asked if anyone had an alternate preferred view.

  18. Stu Clayton says

    “Their scalar implicatures”. Does this mean “their usability in quantitative reasoning” ?

  19. Lars Mathiesen says

    “Scalar implicature” sounds like “what the expression implies about a position on some scale” — in this case, the scale of probability from 0 to 1. Where “not outside” implies “very near the edge” and thus almost zero. (And “possible” doesn’t imply very much at all).

    There are nits to be picked here, since “the realm of possibility” is not defined and it’s not actually said what end of the scale we are near. But if we take the realm to be the half open interval ]0;1] there is in fact only one end where it is possible to go outside (to zero).

    (And of course “not outside” is true of the middle of the scale as well, but that’s one of those “they wouldn’t say it like that if that was the case” things).

  20. David Marjanović says

    I guess it’s “freeze” reinterpreted as “Fritz”, and quite possibly “fried” (of electric components) has also been mixed in.

  21. My subjective feeling is that impossible things, as used in an ordinary speech, usually have probability of about 5% or more. xkcd should draw a cartoon about it and then we can all exchange the link.

  22. Interesting paper about LAS optimization from Google.
    The shear volume of available recorded instances from their speaking/listening tubes must boggle the mind.

  23. “I guess it’s “freeze” reinterpreted as “Fritz”, and quite possibly “fried” (of electric components) has also been mixed in.”

    Yes, imo, though with “fried” as a possible late influence, as electricity apparently doesn’t obtain in the meaning of “on the fritz/friz” in early days.

  24. If you propose 1%-likely etymologies for 69 different words, there’s a 50% chance that one of them is correct.

    Which fact is more helpful for understanding what 1%-likely means than for evaluating proposed etymologies.

  25. PlasticPaddy says

    Your mathematics depends on treating the 69 cases as random trials. Unfortunately someone who proposes 69 new etymologies is likely to be an independent thinker with an overarching theory for all 69, which would invalidate treating them as random trials.

  26. Stu Clayton says

    That’s one of those basic statistical maneuvers that are easy to understand mathematically, but which in concrete cases don’t necessarily help to understand anything. I do understand what “1%-likely” must mean in your particular statement, namely “a possible event with a probability of 1% of occurring”. But a specific proposed etymology of a specific word is not an “event”, nor is its correctness a matter of “probability”, unless you play for Bayes and are invoking different concepts.

    The maneuver is this, crudely put: suppose we have a sample space of 69 possible “correct/incorrect” two-state events, independent of each other. This independence means, by definition, that the probability of any particular combination of the 69 event states is the product of the individual probabilities.

    The probability that (at least) one of the etymologies is correct is 1-W, where W is the probability that all are wrong. For each event, the probability that it is incorrect is 0.99. So W is (0.99)^69, which is just over 0.4998.

Speak Your Mind