Better Phoneme Recognisers Lead to Better Phoneme Posteriorgrams for Search on Speech? An Experimental Analysis

Phoneme posteriorgrams are widely used for speech representation when performing query-by-example search on speech. These posteriorgrams are computed by obtaining the per-frame a posteriori probability of each unit in a phoneme recogniser, regardless the architecture of this phoneme recogniser. It is straightforward to believe that the higher the quality of the phone transcriptions generated by a phoneme recogniser, the higher the quality of its resulting phoneme posteriorgrams; however, to the best of our knowledge, no analysis exist proving this statement. This paper aims at investigating whether there is a correlation between the phone error rate of a recogniser and the maximum term weighted value obtained when performing query-by-example search on speech. Experiments on the Albayzin corpus in Spanish language showed a slight correlation between these two metrics, which suggests that the goodness of phoneme posteriorgram representation is somehow related to phone error rate, but there are other factors that affect their performance in search on speech tasks.