Comparison between Expert Listeners and Continuous Speech Recognizers in Selecting Pronunciation Variants

In this paper, the performance of an automatic transcription tool corpus is by modeling pronunciation variation [2]. is evaluated. The transcription tool is a continuous speech Another way of obtaining models which are less recognizer (CSR) which can be used to select pronunciation contaminated is to train PMs on read speech. It is well known variants (i.e. detect insertions and deletions of phones). The that the extent of variation in spontaneous speech is larger than in performance of the CSR was compared to a reference read speech. So, for read speech there will be fewer mismatches transcription based on the judgments of expert listeners. We between the speech signal and the transcriptions. Thus, it is to be investigated to what extent the degree of agreement between the expected that PMs which are trained on read speech will be less listeners and the CSR was affected by employing various sets of contaminated than those trained on spontaneous speech. phone models (PMs). Overall, the PMs perform more similarly to One can imagine that PMs with varying degrees of the listeners when pronunciation variation is modeled. However, contamination may cause the CSR to select different the various sets of PMs lead to different results for insertion and pronunciation variants. As a consequence, the degree of deletion processes. Furthermore, we found that to a certain agreement between the CSR and the reference transcription may degree, word error rates can be used to predict which set of PMs vary as a function of the PMs employed. The purpose of the to use in the transcription tool. present study is to investigate to what extent the degree of