Tone Recognition of Pahari Language

Abstract

Pahari is an under-resourced, endangered, and undocumented tonal language, spoken in Pakistan Administered State of the Azad Jammu and Kashmir (AJK). Preliminary studies have established the notion, that the Pahari language has three discrete level tones; high, mid, and low. In the current study, tone distribution in monosyllabic words is measured with 45 iterations consisting of 15 high, 15 mid, and 15 low tones, collected from 5 native speakers of Pahari language. An attempt has been made to automatically recognize the phonologically contrastive tones in Pahari language, by using the Random Forest and the Linear Mixed Effect Models with f0 as a preliminary feature along with duration, intensity, F1, F3, and (Cepstral Peak Prominence) CPP. The results showed that the overall accuracy of the Random Forest was higher than the accuracy of the linear mixed effect model. Additionally, the mean f0 played a highly significant role in the prediction of tone while duration, intensity, F1, F3, and CPP played a less significant role.