i2speak

« previous post | next post »

There's a free web-based tool for IPA entry at i2speak.com:


Some people may prefer this to Weston Ruter's excellent IPA keyboard, which has been my IPA-entry method of choice for the past five years. The ability to type mnemonic characters and select from a menu of alternatives, and the availability of specialized keyboards like "Sampa English", ought to make it much easier to enter significant amounts of IPA. I certainly plan to give the i2speak tool a try.

There are a few oddities. For example, the "Vowels" panel shows the traditional IPA vowel quadrilateral, and allows you to click on any of the symbols in order to place them in the output window, as expected. And a few of the symbols have helpful mouse-over popups that give English examples. But the selection is oddly incomplete, and at least one of the choices ([e] exemplified as "met, bed") is wrong, at least in most dialects of English. (See the comments below for some discussion of where this might come from. Perhaps the i2speak implementers might offer a "Setting" panel that would a choice among alternative example-overlays for the vowel display.)

I expect that these few problems will get straightened out as the authors get feedback from users.

Here's a term project for a speech synthesis class: find someone who can perform arbitrary IPA sequences effectively, and create a synthesizer that can pronounce arbitrary IPA strings (at least modulo some sensible phonotactic restrictions). This project poses significant design problems (What synthesis approach to use? How to set up the background database? What sorts of sequences should you attempt to cover?) as well as implementation and testing problems.

There's also a scientific (or perhaps philosophical) dimension. Presumably this "universal IPA synthesizer" would not in fact be at all suitable as a practical synthesizer for all the world's languages. Why not? What would be missing? Could such a system in principle be supplemented with minimal language- or dialect-specific parameters (or even individual-voice parameters) that would bridge the gap between oddly foreign-sounding pronunciations and the level of performance that we expect from a high-quality synthesis system using today's best technology?

Share:



29 Comments »

  1. Simon Kirby said,

    November 28, 2011 @ 8:46 am

    Mark's term project would be more than an interesting intellectual and technical exercise – it would produce a very valuable research tool that I for one would be very keen to have! There are an increasing number of research projects these days that require experimental participants to learn artificial languages which often have very specific phonetic requirements placed on the nonsense words participants have to learn. At the moment, these have to be produced by trained phoneticians, which not every psychology researcher may have access to, and in some cases the number of required items may be prohibitive.

    So, if anyone takes up Mark's suggestion, please do let me know!

    Simon

    [(myl) An interesting (semi-)practical application, which had not occurred to me! I'd suggest that anyone contemplating such a project should get in touch with Simon in order to come up with an approach that would serve such purposes. For example, it might work to limit the system (at least at first) to certain syllable structures, or to impose other restrictions that would make the design and implementation more feasible, while still offering some worthwhile challenges as well as some applications in the artificial-language area that Simon alludes to.]

  2. Martin J Ball said,

    November 28, 2011 @ 8:50 am

    The point about [e] showing examples 'met', 'bed' could suggest the devisers of this handy system learnt their vowel values in Britain, where the vowel in those words would normally be transcribed as /e/, with the vowel in 'mate' always transcribed as a diphthong.

    [(myl) Really? I'm certainly not an expert on British English pronunciation, but as I recall, John Welles gives [ɛ] as the phonetic value of the DRESS vowel in standard British English; and FWIW the Wikipedia page on IPA for English dialects gives [ɛ] as the value for the DRESS vowel in American, Canadian, RP, Irish, Scots, and Welsh English, with [e] given for Australian, New Zealand, and South African. (Though they do supply a footnote saying that this vowel is "Often transcribed /e/ for RP, for example in Collins English Dictionary".)

    The choice of [ɛ] corresponds to my impressions — certainly when I listen to BBC news readers and the like, I don't hear their DRESS vowels as being nearly as high as [e].

    I don't know of any equivalent to Peterson/Barney or Hillenbrand et al. for RP — assuming that something comparable hasn't already been done, a worthwhile Breakfast Experiment™ would be to look at the formant-space vowel measurements of some BBC news readers. I'd be very surprised to find that the trajectory of their FACE vowels isn't significantly fronter and higher than the trajectory of their DRESS vowels — and I'd predict that the overall position of their DRESS vowels would not be very different from the corresponding position in their American-English counterparts.]

    [To clarify: there is no question about whether the DRESS vowel in British English is [e] or [ɛ]: Mark is right, it is definitely [ɛ]. But what Martin is saying is also right: British IPA phoneticians, when writing a broad transcription, have often written that vowel with [e], avoiding a special symbol not on the roman keyboard, because the BET vs. BAIT distinction can be represented with [bet] vs. [beit]. Although British English has [ɛ], it doesn't have a distinction between pure vowels [e] and [ɛ], so the IPA practice of using the nearest ordinary roman letter that doesn't make for any ambiguity can be followed. Alles klar? —GKP]

  3. Paul said,

    November 28, 2011 @ 9:52 am

    myl, you may have overlooked Martin's careful distinction between [e] and /e/. British speakers do indeed generally have something phonetically closer to [ɛ] than [e] in DRESS, but a major tradition in transcription has the phoneme written as /e/ (which follows the style of using roman characters where possible). The use of IPA-type symbols for phonological units (ie phonemes) as well as phonetic transcription is a never-ending source of confusion, and presumably it has also caught out the programmers of this keyboard – they haven't decided whether what they're doing is phonetics or phonology. As a further example, I used to find it hard to teach students the proper phonetic value of [ʌ] alongside their seeing the same symbol used phonemically for the STRUT vowel (which was closer to [ə] or [ɐ] for most of them).

    In any case, it seems a bit random which vowels have keywords with them – and the thing could be improved by removing the example words completely. If users don't know enough phonetics, the example words can be misleading (witness this discussion); and if they do know enough phonetics, the example words are superfluous.

  4. Paul said,

    November 28, 2011 @ 9:54 am

    PS I rather like this: http://people.w3.org/rishida/scripts/pickers/ipa/

  5. NW said,

    November 28, 2011 @ 10:33 am

    The value of RP /e/ has also shifted in recent decades. It used to be midway between the two cardinals, as it still is for some of us, but has moved down to cardinal mid-low, or even further. Young women seem to have the most shifting: I overheard a 'yes' the other day that was almost [jas], with close to cardinal [a], and this is not unusual.

  6. Chris said,

    November 28, 2011 @ 10:44 am

    Frikkin thing crashed my Droid! (Droid X2, Firefox)

  7. Joe said,

    November 28, 2011 @ 10:54 am

    Probably no need to add this after GKP's comment, but here is Wells himself on the issue:

    http://phonetic-blog.blogspot.com/2009/03/e-and.html

  8. dw said,

    November 28, 2011 @ 3:17 pm

    I still prefer the keyboard at http://ipa.typeit.org/full/ because it has keyboard shortcuts.

    For example, CTRL + T will give θ

  9. Antariksh Bothale said,

    November 28, 2011 @ 3:55 pm

    I have a question about the [e] v/s [ɛ] issue. Aren't we essentially undermining the utility of the IPA by again introducing ambiguity in phonetic transcription? IMHO, if a person has learnt the IPA and has some understanding of what sound each symbol represents, he should be able to 'know' (find out) the pronunciation of any word directly by seeing its phonetic transcription. If you need native knowledge to find out how a word is pronounced, won't the IPA loses its value? You could just have spelling instead. Any native speaker, irrespective of whether they know the IPA, would tell you how DRESS is pronounced.

    Introducing multiple standards there is again sending us back to shaky ground—a ESL speaker, for instance, can never be sure how to pronounce it because he won't know what sound a particular symbol (which should technically have been unambiguous) is supposed to represent. And 'pronunciation keys' given in the beginning of dictionaries are often not very helpful for ESL speakers, because saying [e] as in dress still doesn't give you any real idea as to what the sound is unless you can conjure a native speaker and make them pronounce the word. At least with the IPA, you can verify, independently, the sound that [ɛ] represents, and try to pronounce it.

    As for making the transcription easy by having as many roman letters as possible, doesn't it increase the chances of inaccuracy? For instance, Hindi doesn't have the [t] sound. Instead, it has a retroflex stop [ʈ] and a dental stop [t̪]. Now, one might argue for using [t] to represent the dental stop instead of [t̪] in order to avoid writing the diacritic and to have a simple roman letter transcription, but I think it contributes to ambiguity about its pronunciation.

    Note: All the points I have made are with respect to non-native speakers trying to find out how a word is pronounced in the language.

  10. Rohan Dharwadkar said,

    November 28, 2011 @ 5:36 pm

    @Antariksh: Even with phonetic transcription, there can be various degrees of accuracy depending on purpose and target audience. For instance, the phoneme /r/ in RP is usually [ɻ] or [ɹ] phonetically. However, it is often labialised, and this additional bit of phonetic information may or may not be tacked on to the phonetic transcription, depending on the aforementioned factors.

    In addition, the transcriptions provided in most dictionaries are *phonemic* transcriptions. So in effect, for a foreign language learner to pronounce a given word 'correctly', the dictionary transcription needs to be supplemented by (at least some) knowledge of the phonology of the language in question. Thus, to pronounce the word transcribed as [pet] in an English dictionary, the ESL learner would also need to be aware of the fact that syllable-initial voiceless stops are aspirated in RP.

    The perceived loss of the IPA's value has, I'm sure, incited numerous phoneticians to devise fixes. One such phonetician that I'm aware of is Luciano Canepari, who has apparently assigned independent, diacritic-free symbols to all phones imaginable, and then analysed several languages using his system. Check it out here: http://venus.unive.it/canipa/dokuwiki/doku.php?id=en:start

  11. LDavidH said,

    November 28, 2011 @ 5:38 pm

    @Antariksh Bothale: As an ESL speaker, I see what you're saying – but it's not entirely true. If I read that MANY, despite being spelt with an A, contains the same vowel as DRESS, I know how to pronounce it because I know the basic vowel sounds of English, and I will say it the same way that I say DRESS – whether my version is more of an [e] or an [ɛ] sound.

    On the other hand, being told that a Hindi word contains the sound [t̪] doesn't actually help me, since reading a phonetic description of a sound, however accurate and exact, is only helpful to people who understand the technical terminology. Unless a native Hindi speaker lets me hear the difference between [t̪] and [ʈ], all I will be able to do is pronounce them the way I normally pronounce T, knowing that it's probably wrong but not being able to correct myself. The IPA is helpful if you know (by hearing) the actual sounds the symbols refer to; if you haven't heard them, the IPA doesn't really help very much.
    My native language, Swedish, has the sound [ʉ:]; but very few people will know what that means. If I say that it is basically the same as the vowel in Southern BrE FOOD, at least my English wife knows what I mean!

  12. Rohan Dharwadkar said,

    November 28, 2011 @ 5:38 pm

    Sorry. Should've been /pet/.

  13. Rohan Dharwadkar said,

    November 28, 2011 @ 5:39 pm

    Also should've been 'syllable-initial STRESSED voiceless stops'. Damn!

  14. Sandy Nicholson said,

    November 28, 2011 @ 5:42 pm

    Regarding GKP’s clarification above, it should be noted that the distinction between [e] and [ɛ] is made in most (all?) Scottish dialects – so his comments don’t apply to _all_ varieties of British English. I have to say that before I’d studied any phonetics, I was misled by dictionaries like the OED into thinking that my [e] really was an [ɛɪ] glide – and almost convinced myself that I could hear it (though I felt it was actually nearer to [ɛi] if it _had_ to be a diphthong). Other supposed diphthongs were even more of a mystery to me, and I just assumed that there was something very complicated going on, perhaps known only to a hallowed inner circle of lexicographers. (I’m not sure whether to /bɛt/ on Geoff taking the /bet/, though.)

  15. Dominik Lukes said,

    November 28, 2011 @ 6:38 pm

    This is great and I suspect I will be using the i2Speak SAMPA implementation as my main way of typing IPA (as a non-phonetician, my needs are simple). But I also like the limited http://ipa.typeit.org as a tool to recommend to students who are training to become teachers and only need to transcribe a few words here and there to practice.

  16. dw said,

    November 28, 2011 @ 7:07 pm

    @Dominik Lukes:

    Check out the http://ipa.typeit.org/full/ as well: not "limited" in the same way as http://ipa.typeit.org

  17. dw said,

    November 28, 2011 @ 7:17 pm

    @Antariksh Bothale:

    It's not a question of "different standards", but of broad versus narrow transcriptions.

    See, for example, the section on "American English" in the IPA Handbook, available (I hope) via Google Books here.

    On page 44, there are two transcriptions given of the same passage. The first, a broad transcription", is to be interpreted with the aid of the "conventions" specified on page 43. The second is a narrow transcription, which can be interpreted simply by knowing the definitions of the IPA symbols.

  18. Adam said,

    November 29, 2011 @ 4:29 am

    FWIW, the OED has:

    bet /bɛt/
    bait /beɪt/

    and the following explanations in the pop-ups:

    IPA Sounds like
    ɛ e as in pet, ten
    eɪ ay as in bay

  19. LDavidH said,

    November 29, 2011 @ 6:08 am

    In addition to my previous comment: the IPA is of course useful to indicate the difference between, say, BrE and AmE pronunciations, or BrE and Australian. But you still need to have an aural idea of the sounds described.

    And also, whose pronunciation should be seen as the most correct, since the IPA has to be absolutely accurate? Should the IPA indicate the slight [t]-sound heard in many BrE speakers' version of PRINCE, MINCE etc? And what does the IPA do about the "intrusive /r/" in "law and…"? Should that be indicated as standard, or left for us ESL speakers to discover for ourselves?
    All in all, I find the "/e/ as in pet" more helpful than the IPA (but maybe that's because I live in the UK and can just ask my wife…)

  20. Leonardo Boiko said,

    November 29, 2011 @ 7:57 am

    @LDavidH: Most times you can adequately approximate the sound yourself, by taking the IPA descriptions as instructions. I have no contact at all with Hindi, but I can make a [t̪] myself by starting with a [t] and then making it more dental. If I was learning Hindi, I’d pronounce this a few times alternating with a regular alveolar [t], trying to understand the distinction, and only then confirm it with audio examples from the Internet.

    For me, such meta-linguistic awareness helps a lot when learning new phonemes. I would never have noticed that the Japanese sound in "shi" is actually a [ɕ] and not [ʃ], or indeed that [ɕ] even exist, if I hadn’t looked up IPA descriptions of Japanese and tried to reproduce the articulatory characteristics of [ɕ]. And I had been listening to native speakers for years. In the same way, it was conscious and active study of the IPA descriptions that enabled me to finally perceive the Mandarin distinctions [ʂ] [tʂ] [tʂʰ]/[ɕ] [tɕ] [tɕʰ] when listening to natives (“ok, so it’s like an [s] but retroflex, so if I pull my tongue like this… [ʂi] [ʂi] [si] [si] [ʂi] oh, so _that_ is what this sound was supposed to be like… why is everyone in the cafe looking at me?”)

  21. Leonardo Boiko said,

    November 29, 2011 @ 8:01 am

    @Adam: On the other hand, the OALD has [bet] [beɪt], which confused the hell out of me back then. I recall thinking that “breakfast” must have been “supposed” to be pronounced with the same [e] of “break”, despite the fact that I could distinctly hear its [ɛ]. And it’s a learner’s dictionary…

  22. LDavidH said,

    November 29, 2011 @ 10:19 am

    @Leonardo: Well, in theory I guess you're right – although it still means having to learn what "retroflex", "alveolar", etc mean, never mind trying to bend your tongue and mouth muscles into the right position. Most ESL speakers are not linguists, only non-native speakers wanting to communicate, and don't have the energy or the interest to learn all the technical terms. I still think "MANY rhymes with PENNY" is more helpful to 96% of learners, than saying that the "a" in MANY is whatever it might be in phonetic description. (And why is it different in MANIFOLD??)

  23. Antariksh Bothale said,

    November 29, 2011 @ 12:02 pm

    I think we are confusing two things here: I am not saying that that IPA transcriptions of English words should be the sole way for ESL learners to find out the pronunciation of English words. However, assuming they do know how PENNY sounds, I would rather tell them [ɛ] as Penny rather than [e] as in Penny.

    LDavidH, contrary to what you claimed, it is much easier to convey what the Swedish sound is by telling that its [ʉ:]. Saying that it is the same as a sound in some other language is much more helpful, yes, but only if you know the pronunciation of the other word accurately. I am perfectly fluent in English, but I cannot possibly be sure as to what the South BrE pronunciation of FOOD is, but I can instantly pronounce [ʉ:] simply because I know it's a long high central rounded vowel. If Swedish books transcribed it as a simple [u:], I would never even become aware of the fact that it has less backness than the [u:]. Those who don't know phonetics can still be told that it's like the South BrE FOOD, but I see no reason to try to simplify it (and hence render it inaccurate) for those who do.

    Also, I have a fundamental problem with trying to teach sounds by asking people to relate them blindly to their mother tongue. I am learning Mandarin right now, and I can see what disaster results when most students try to pronounce Mandarin by basing it on Hindi sounds. If you look at the IPA-chart for both languages, you will see that the most of the Chinese fricatives and affricates aren't there in Hindi, but you could still attempt to teach them by approximating them to Hindi sounds. This is OK for a learner who is aware of the differences and is only using his language as a base, but it is bad for most other students who then become deaf to the actual differences in vowel and consonant quality and simply keep repeating Hindi sounds. Almost all phonemic differences are lost and it all collapses into two Hindi sounds.

  24. LDavidH said,

    November 29, 2011 @ 4:09 pm

    @Antariksh: "I have a fundamental problem with trying to teach sounds by asking people to relate them blindly to their mother tongue."

    Yes, in theory, so do I – but in actual practice, for most non-linguists, that is the best they will ever do anyway. My English wife knows how the long Swedish vowels Y [yː], O [uː], and U [ʉː] sound when I say them; she still can't actually reproduce the first two of them. We used to live in Albania; the Albanian language has two CH-sounds, the voiceless palatal plosive [c] and the voiceless postalveolar affricate [tʃ]. My wife was never able to hear the difference, never mind producing the first one; giving her the exact phonetic description would have been completely pointless. I could just about hear the difference, and I pronounced them differently; but only because I asked native speakers to pronounce them for me. Again, I can't "flesh out" the description into an actual sound. I'm pleased you can, it must be very helpful! But for most of us, an approximation is the best we'll achieve (as witnessed by the fact that I still speak English with a slight Swedish accent, despite being surrounded by native speakers!).

  25. Ellen K. said,

    November 29, 2011 @ 6:11 pm

    One brief thought. Different people learn differently. What one person finds to be the best method for learning the sounds of a foreign language may not be so helpful to another person.

  26. dw said,

    November 29, 2011 @ 8:12 pm

    @Antariksh Bothale:

    I have a fundamental problem with trying to teach sounds by asking people to relate them blindly to their mother tongue

    Even if learners are not taught to do this, doesn't this happen inevitably, given the way the human mind works? The more gifted linguists will go on to acquire a grasp of the phonology of the new language: the less gifted will never get beyond a mapping of phonemes from new language to old.

  27. Antariksh Bothale said,

    November 29, 2011 @ 9:52 pm

    @LDavidH and dw:

    I agree with you on the problems you mention, but I was just wondering whether it would be a better pedagogical practice to help students understand (in brief) the phonetics involved in the whole process. For instance, I often try to teach friends how to pronounce sounds—especially vowels—of foreign languages, and I have found that a mix of phonetics and mother tongue is better than letting them pronounce it just like that.

    For instance, the high front rounded vowel [y], the rounded counterpart of [i] is not there in Hindi phonology. Now, when learning French, Indians either say [u] (since they have only pronounced back rounded vowels before) or [yu], which is significantly better and which you can find in Hindi word. The spelling 'u' doesn't help either. In this case then, I prefer to teach them this sound by first pointing out the rounding they do for [u], then making them pronounce [u] and slowly try to move their tongue forward. In simple words, "pucker your lips like you want to say oo and then try to say ee instead". It takes a few tries, but I have gotten good results in general. I feel this is better than simply letting them pronounce [yu].

  28. Antariksh Bothale said,

    November 29, 2011 @ 9:56 pm

    (sorry, forgot to add this)
    This technique has worked brilliantly well for teaching them the pronunciation of [ʒ], which doesn't exist in Hindi. I first point out the voicing they do in other voiced-unvoiced pairs, then ask them to try and 'vibrate' their [ʃ]. This has worked almost every time I've tried it. The other technique, which is 'speak as it sounds', has almost never worked. They just end up saying one of [ʃ] [z] or [dʒ].

  29. LDavidH said,

    November 30, 2011 @ 3:30 am

    @Antariksh: Yes, that's obviously the best approach – but that's in a situation where a teacher can make the correct sound _and_ explain to others how they do it. My criticism of IPA is based on trying to work out how to say something from only a written source, no native (or skilled non-native) around to explain. And sadly, very rarely are native speakers able to explain how they make a certain sound.
    In principle, I agree that you shouldn't just accept "mother tongue sounds" if it can be helped; I just have too much experience of cases where that was better than having people so tongue-tied that they wouldn't speak at all… (which, by the way, applies to grammar as well).
    In addition: if IPA [y] refers both to French U and to Swedish Y, it's unreliable anyway; I speak good French and can testify that they are not the same sound. The Swedish Y is much closer to an I (although clearly distinguishable); French U is somewhere between the Swedish U and Y; or, Swedish U is somewhere between French OU and U (but nearer to U).

RSS feed for comments on this post · TrackBack URI

Leave a Comment