Biblical Hebrew Forum

[James Christian <jc.bhebrew AT googlemail.com>]

Hi,

good question. The way we moving from phone recognition to graphic form is
typically handled in state of the art speech recognition software is by
recourse to language modelling. That is to say we take a massive written
corpus of the target language (e.g. English) and make n-gram counts and
therefor are able to build a model of the most likely sequences of words.
Using this kind of information we are able to make statistical guesses about
how the phonetic representation should map to the perceived graphic
representation. Well trained state of the art systems used in the industry
(like subtitle generators for live news broadcasts) typically get over 97%
accuracy using the WER (word error rate) metric. Compared with human
performance this is quite competitive and so we have good reason to believe
we are working in the right direction.

James Christian

2010/1/24 s.a.breyer <s.a.breyer AT gmail.com>

> It seems to me that James and Yitzhak are talking about two different
> things.
>
> James is concerned with the acoustic fact, and his account of that is
> accurate. A significant piece of my job involves editing audio and video
> recordings, correcting misprounounced words, cutting in different words
> and the like; my software is nowhere near so informative as that I
> presume James works with, but even so there can be no question: a fully
> articulated English consonant is tied to the following vocalization.
>
> Yitzhak, however, is concerned with the linguistic fact, and his account
> of that is equally accurate.
>
> The two accounts are to some extent mediated by the physiological fact
> that the central consonant cluster of the phrase in question does not
> stand by itself: it requires an initial movement of the apex of the
> tongue to the alveolar ridge and consequent closure of the vocal stream;
> that movement, it seems to me, is "perceived" by the speaker and hearer
> as a relization of the consonant which closes the word or "syllable". In
> normal speech, in fact, that's all you get even if the consonant closes
> an utterance; if you ask someone "Did you do this?" and he answers "I
> did." all you're normally going to hear at the end is the closure, not a
> fully enunciated and voiced "d". (Significantly, this deletion is less
> marked with unvoiced consonants; a voicing probably implies a subsequent
> vowel.)
>
> (This is, by the way, not "sloppiness", as people are fond of
> suggesting; the full enunciation the schoolmarms propound is an abnormal
> articulation. To be sure, stage actors, particularly classical actors,
> are trained to speak this way, but that's because the stage actor has to
> deal with abnormal acoustic situations in which making the linguistic
> substratum intelligible takes precedence over making the speech stream
> sound "natural".)
>
> The *interesting* question is how people parse the acoustic fact into
> the linguistic fact - which is what I'd really like to see James write
> about instead of all this who-shot-John about an entity which is of
> little use outside prosodic analysis.
>
> Stoney Breyer
> Writer/Touchwood Creative
>
>