Recognition of Spoken Languages from Acoustic Speech Signals Using Fourier Parameters

摘要

Spoken language identification (LID) or spoken language recognition (LR) is defined as the process of recognizing the language from speech utterance. In this paper, a new Fourier parameter (FP) model is proposed for the task of speaker-independent spoken language recognition. The performance of the proposed FP features is analyzed and compared with the legacy mel-frequency cepstral coefficient (MFCC) features. Two multilingual databases, namely Indian Institute of Technology Kharagpur Multilingual Indian Language Speech Corpus (IITKGP-MLILSC) and Oriental Language Recognition Speech Corpus (AP18-OLR), are used to extract FP and MFCC features. Spoken LID/LR models are developed with the extracted FP and MFCC features using three classifiers, namely support vector machines, feed-forward artificial neural networks, and deep neural networks. Experimental results show that the proposed FP features can effectively recognize different languages from speech signals. It can also be observed that the recognition performance is significantly improved when compared to MFCC features. Further, the recognition performance is enhanced when MFCC and FP features are combined.