An efficient combination of acoustic and supra-segmental informations in a speech recognition system

A fast and reliable rate of speech detector

“…It has been argued [2,3] that phone rate is more suited than syllable or word rate. By normalizing the phone durations with respect to the phone specific expected durations [4,5] and variances [2], a normalized phone rate can be obtained that is very effective in differentiating utterance rates. However, this requires phonetic segmentation and classification information that is bound to be provided by the recognition process.…”

“…However, this requires phonetic segmentation and classification information that is bound to be provided by the recognition process. Therefore, normalized phone rates can only be calculated during [4] or after [2,5] the recognition search.…”

“…On the other hand, this technique has the disadvantage that the recognition can only start after the completion of the utterance. This approach was also used in the multi-pass search algorithm reported in [4]. The mean syllabic duration was measured on an entire first recognition hypothesis and subsequently used to adapt the subword unit durational characteristics which are used in a second recognition pass.…”

“…Typical choices for the initial estimate are the statistical mean of the ROS and the final estimate of the previous sentence. This approach was used in [4], where a speaking rate factor is deduced for each recognition hypothesis by means of a Kalman filter. In that case, the ROS calculation is integrated in the search and at each time instant, a normalized ROS estimate is calculated for each recognition hypotheses.…”

“…It has been shown that compensation techniques can reduce the errors for fast speech in HMM [2,3,4,5] as well as in hybrid HMM/MLP [6] recognition systems. However, these techniques require a reliable ROS detector.…”