Deep Neural Networks for Spoken Language Identification in Short Utterances

This work presents the elements of language identification (LID) in small segments created using short duration utterances. For low-resourced languages availability of data itself is a challenge. The paper tries to apply DNN for low resourced language. This paper presents a feed-forward deep neural network (FF-DNN) for language identification using acoustic features of short-time utterances. Two network topologies for DNN have been checked for their performance in LID task. The obtained findings of the experiments are compared to a well-established technique based on i-vector system. This i-vector system uses MFCC-SDC to represent speech feature that represent the acoustic characteristics and the back end is implemented using support vector machine (SVM) that serves as a classifier. These mechanisms were put in place to help with identification of Hindi and Punjabi, two widely spoken Indian languages. The speech utterances are divided into short segments of 5 s, 10 s, 20 s and 35-s duration. The system’s efficiency is measured in EER (%) and for short time segments, a relative improvement of 3% is achieved by the DNN system, whereas the average error rate overall the utterances was decreased by 2% using DNN.