The Sonorant Feature in Language Identification

Abstract

This project is aimed at taking the first steps towards the creation of a "landscape map" by testing whether or not the sonorant feature is a relevant feature in language identification. Sonorant sounds are those with free and easy voicing including glides, nasals and vowels. The statistic of the sonorant we use is the average sonorant value over the speech sample. Using techniques of signal processing and pattern recognition, we use three experimental setups. We first build a system to identify sonorant patterns. Our hypothesis is that the sonorant feature is relevant in language identification. If the classification by a system using the sonorant feature is more accurate than the classification by a similar system without using the sonorant feature when both systems are trained and tested on the same data sets, we may conclude our hypothesis is true. To test this hypothesis, we will use an experimental setup which attempts to classify languages without using the sonorant information using a hidden Markov model and a second level support vector machine-based recognizer, and another setup which will augment this feature set with sonorant information. If we find a significant improvement in accuracy by using the sonorant information, we will conclude that the sonorant patterns are a relevant feature for the language map.