tag:blogger.com,1999:blog-1268266350438382990.post7609492361539726804..comments2023-09-29T04:08:18.583-07:00Comments on Spoken Language Processing: Are segmental durations normally distributed?Anonymoushttp://www.blogger.com/profile/06723264194529654613noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-1268266350438382990.post-44334314033443971672010-11-03T10:09:44.308-07:002010-11-03T10:09:44.308-07:00The problem is that phone durations (in the Boston...The problem is that phone durations (in the Boston University Radio News Corpus) aren't all that log-normally distributed. When i look at the KL divergence between the MLE Gaussian and a histogram of durations or log durations, the KL-div is lower when a normal distribution is used, rather than log-normal. However, one difficulty in measuring this is that the BURNC forced alignment data has a very coarse granularity with durations quantized to 10ms frames. This leads to a sparse histogram in log-space. I'll take a look at this on TIMIT which has a much finer (and manually constructed) annotation, and follow up about whether phone durations are more log-normal or half-normal.Anonymoushttps://www.blogger.com/profile/06723264194529654613noreply@blogger.comtag:blogger.com,1999:blog-1268266350438382990.post-53093955858479656112010-11-03T04:08:28.773-07:002010-11-03T04:08:28.773-07:00I have used log-normal distribution to model segme...I have used log-normal distribution to model segmental durations of phonemes.Anonymoushttps://www.blogger.com/profile/04297766304935234782noreply@blogger.com