An investigation on prenasal merger in Southern American English through automatic speech recognition

Commonly used in Automatic Speech Recognition programs, forced alignment is the process of generating phone-level timestamped transcriptions based on orthographic transcriptions. In this work, a modified pronunciation dictionary for forced alignment is used to identify the prenasal merger (a.k.a. pin-pen merger) in Southern US English. We hypothesize that the modification will increase the acoustic separation between the prenasal allophones of IH and EH. This is borne out in our experiments performed on the Digital Archive of SouthernSpeech audio corpus. We compare vowel formant values before and after the modification, including the separation between vowel clusters as measured by Pillai scores and Euclidean distances between vowel centroids. K-means clustering on vowel formant values is used to show our modification yields better phonetic transcriptions. We also use Kullback-LeiblerDivergence to show the trained acoustic models for the prenasal allophones of IH and EH differ the most in their final portions.