Language Identification using Stacked Convolutional Neural Network (SCNN)

The Automatic Speech Reorganization (ASR) system has a wide range of industrial and practical applications. The Language Identification (LID) is the first step of the ASR and utilized in various applications. LID system is used to classify the spoken language from an audio signal. The automatic language detection provides correctly parsed speech utterance and grammar rules to the speech recognition process. In this work, we proposed an image domain LID system instead of an audio domain. A Stacked Convolutional Neural Network (SCNN) is used on the extracted spectrogram images of an audio signal. The features are extracted from the image domain signal and classified based on their sampling rate and frequency. The feature extraction and language detection are performed by the SCNN method. The performance evaluation of the proposed method is done based on accuracy, precision, and recall parameters. The language identified by SCNN for the EU speech dataset is more accurate than the CNN-RNN approach. The SCNN based LID model can be used for the large scale dataset and implemented on the long uttered audio signal.