Is a perfect Siri possible?

We've all experienced the frustration of having a computer speech system fail gloriously to understand what we're tyring to tell it. It just so happens that when I use Siri, shy seemingly gives up as soon as it's confused about what I'm saying, and then converts all that it did "hear" – in other words, actively record and process – into the most probable transcription. This transcription, without the context of the rest of my utterance, is almost always and entirely wrong. But it seems that this error arises also in some sense from a "mishearing" and thus a mis-transcribing of what was said. Although this is most probably not how Siri truly functions, we can imagine, as she types on on the screen what she hears us say, that she is a mapping from sound to writing.

In the Kenstowicz reading, we are told of the distinction between allophones and phonemes. This distinction explains, at least in part, the reason for there being such a breadth of ways any given "sound" or letter in English (or potentially any language) is actually pronounced. The logical unit we learn, the phoneme, is not always manifested as the same allophone. While Kenstowicz discusses rules that govern these allophone-phoneme transformations, we can see that this distinction provides a challenge for any system that attempts to convert sound to writing of the same language.

While not referenced directly by Gussenhoven (except in multiple redirections of the reader to the table on page eighteen), the IPA – International Phonetic Alphabet – is, supposedly, a system that allows us to represent sounds created by human speech directly as writing. In essence, there should be a unique mapping from a sound to a sequence of symbols representing it. (This is also hinted at in Kenstoqicz, who mentions the choice of phonemes a language makes from the posited Universal Grammar, which would then serve as a set of sounds.) If the IPA can truly function is the way, abstracting a program such as Siri into a mapping from sound to IPA, and then from IPA to possible language-specific interpretations would potentially function very well. Going from IPA to possible "real writing" in a language, while a hard an interesting problem, does not serve any immediate challenge to this idea; the problem with this approach lies in the assumption of the existence of a way to go from sound to IPA (or really any symbolic representation). No human "speaks" IPA in a native sense, and thus even IPA has the potential to not be unambiguous, with people having different interpretations of symbols, even if the symbols are not supposed to allow for them. This question shows only whether IPA is the correct system to target the transcription of sound.

The real question is does such a system exist. If it does, maybe a perfect Siri is not that far off.