Language Recognition By Convolutional NeuralNetworks

Speech recognition representing acommunication between computers and human as a sub fieldof computational linguistics or natural language processing hasa long history. Automatic Speech Recognition (ASR), Text toSpeech (TTS), speech to text, Continuous Speech Recognition(CSR), and interactive voice response systems are differentapproaches to solving problems in this area. The performanceimprovement is partially attributed to the ability of the DeepNeural Network (DNN) to model complex correlations inspeech features. In this paper, unlike the use of conventionalmodel for sequential data like voice that employs RecurrentNeural Network (RNNs) with the emergence of differentarchitectures in deep networks and good performance ofConventional Neural Networks (CNNs) in image processingand feature extraction, the application of CNNs was developedin other domains. It was shown that prosodic features forPersian language could be extracted via CNNs forsegmentation and labeling speech for short texts. By using ۱۲۸and ۲۰۰ filters for CNN and special architectures, ۱۹.۴۶ errorin detection rate and better time consumption than RNNs wereobtained. In addition, CNN simplifies the learning procedure.Experimental results show that CNN networks can be a goodfeature extractor for speech recognition in various languages.

Keywords:

Speech recognition , Speech Segmentation , Convolutional neural networks , Persian Language , Artificialintelligence.