Speech Recognition Improvements

To recognize speech input, we used the WAMI (Web

-

Accessible Multimodal Interface) software [6], a

framework that allows audio to be captured at the web

page and transmitted to the SUMMIT speech recognizer

[5] running remotely.

During initial user testing

, we

observed

that false

negative speech recognition errors

(i.e. user spoke

correctly but speech recognizer misrecognized)

were

most common and significantly

inhibited the flow and

enjo

yment of the game. F

alse positive errors were

rarer

because users tend

ed to speak only when they were

fairly confident in their answer

. Because the target

answer was

revealed whenever

the

user

succeeded,

users appeared amused rather than misdirected by

false positives. To decrease the likelihood of false

negatives, we

add wo

rds to

the recognizer’s vocabulary

dynamically as words are introduced in the game, and

furthermore expand the set of accepted responses to

include the top

three

speech recognition

hypotheses

once the vocabulary size reaches a

sufficiently large

threshold.

Thus, if the correct word a

ppears anywhere

in the top few

words, it is selected as the user’s

utterance.