tlhIngan-Hol Archive: Wed May 15 12:40:43 1996

Back to archive top level

To this year's listing

[Date Prev][Date Next][Thread Prev][Thread Next]

Klingon CD-ROM

I can answer some of the questions about the speech recognition on 
the Klingon CD-ROM. We built the recognition here at Dragon 
Systems (ghargh'a' patmey malja'ghom). I designed part of it, 
keeping in touch with Marc Okrand, Touchscreen Media Group (who 
put the CD-ROM together), and Simon & Schuster Interactive (the 
publisher).

Normally, when we build a recognizer, we collect a sizeable body 
of pronunciations of many different words by many native speakers, 
and we then build computer models of each phoneme (elementary 
speech sound; in Klingon these correspond to the letters: a, b, 
ch, ... w, y, ') in many different contexts. Then we build the 
pronunciation models for the vocabulary from these phoneme models.

In this case we did not have the time or resources to do that. 
Furthermore, this recognizer had to be able to handle 
MISpronunciations as well as correct pronunciations. So we tried 
to guess how people would mispronounce the words in the 
vocabulary, and recorded speakers (9 male, 7 female) pronouncing 
each word correctly and in each expected incorrect way, and we 
built whole-word models from each pronunciation. This method 
requires less data, but it is much less flexible and gives poorer 
recognition performance. Among other problems, it's much more 
dependent on the particular voices and pronunciations of the 
recording speakers, which may be responsible for some of the 
problems DTrimboli reports. 

Only after all the speakers were recorded did we receive the audio 
files of Robert O'Reilly's ("Gawron"'s) pronunciation of the 
words. Now, Okrand coached him all through the recording session, 
and on the audio files I could sometimes hear MO in the 
background, correcting or encouraging him. (That was edited out of 
the product, of course.) In my conversations with Okrand, he has 
nothing but praise for O'Reilly's efforts at correct 
pronunciation. 

But in the tutorial, ROR often draws out the pronunciation, saying 
a word emotionally as in the total immersion script; and sometimes 
he's just plain wrong. Okrand and I agreed that it was preferable 
in this game to give "Qapla'!" to a mispronunciation than to 
correct a good pronunciation, if we had to choose; and in 
particular that an imitation of ROR should always be counted as 
right. So wherever ROR said a word in a way matching a recorded 
mispronunciation, I modified one of the control files so that 
mispronunciation would be considered correct. But when he 
mispronounced in a totally different way, there was nothing I 
could do. I've heard, though, that Okrand's recorded corrections 
then guide them to a correct pronunciation. 

Another problem we could not avoid was the variety of hardware. 
When people talk on the phone, you know that their voices sound 
different depending on the phone they're talking from. Well, 
different computer microphones have a similar effect, and so do 
different sound boards and different software settings for them; 
and we had no control over these. For our own products we test and 
certify many kinds of mike and board. We can recommend certain 
equipment, and sometimes we have to say "mike X will produce poor 
results with our system; we advise against using it". We can even 
build the product to adjust itself in certain ways. But we 
couldn't do any of those things here. No computer yet built can 
match the human ear and brain in compensating for these channel 
variations, and we weren't even able to try. 

I think I've put enough time into this for right now, and I hope 
my answer explains a few things about the Klingon Language Lab: 
      <<qay'ghachmey'e' boyaj boneHbogh vIQIj 'e' vInIDta'>> 

      marqem, tlhIngan veQbeq la'Hom -- Heghbej ghIHmoHwI'pu'! 
     Subcommander Markemm, 
            Klingon Sanitation Corps -- Death to Litterbugs! 

                Mark A. Mandel : [email protected] 
    Dragon Systems, Inc. : speech recognition : +1 617 965-5200 
 320 Nevada St. : Newton, MA 02160, USA : http://www.dragonsys.com/

  +-------------------------------------------------------------+ 
  | KLINGON PAGE! http://www.dragonsys.com/klingon/klingon.html | 
  +-------------------------------------------------------------+

Back to archive top level