Corpus Design Challenges and Analysis of Acoustic Information for Automatic Dialect Classification

The primary challenge across the field of automatic speech communication relies upon the understanding of creation of such acoustic models that are capable of representing distinct differences throughout the spoken languages. The design of such systems is undoubtedly affected by the parameters such as an Individual’s age, gender, speaking style, dialectal variability, recording environment etc. This work investigates on the various challenges faced during designing an enriching corpus and analyzing the dialectal variations by tracking the variance of different acoustic parameters such as pitch contour, formant frequency, and intensity for different toneme sounds. The article also makes an attempt to discuss the process and problem of tone assimilation (tone sandhi) in di/multi-syllable words of Punjabi language where the occurrence of more than one tone is being anticipated. The pitch variations of the input signal are performed using PRATT toolkit, to establish a relation between the pitch variations across the various tonemes. The model is further evaluated using the ANOVA test for the comparative classification of dialects based on acoustic features. The entire work is based upon the four dialects of Punjabi language; classified as the 10th most spoken language across the world, with over 105 million native speakers.