Pause skip

In our survey of erroneous results (section 4.2), we found that the wrong sentences included $/pause/$ in the speech data. We concluded that such $/pause/$ causes errors because acoustic parameters and word trigram models do not correspond. To counter this, we modified the recognition algorithm, based on the following techniques. As $/pause/$ is considered to be one word, it can be recognized, and thus, the trigram probability can be calculated to skip it. For example, when we see that w1 = $/pause/$ in our calculation of $P(w0\vert w2,w1)$, we calculate $P(w0\vert w3,w2)$ instead. And if we calculate $P(w0\vert w2,w1)$, and $w0 = /pause/$, we set $P(w0\vert w2,w1)$ to $1$.

This algorithm raised the sentence recognition rate to 71.6% as opposed to 66.7% with no such procedure for the SD case. Furthermore, a great improvement was found for the SI case (61.7% as opposed to 0.0% with no such procedure).