HindiSpeech-Net: a deep learning based robust automatic speech recognition system for Hindi language

Automatic Speech Recognition (ASR) has become one of the major research areas over the past decade and gained a lot of interest. Their system implementation, adaptation to different languages and robustness in the performance are still some of the major challenges. Hindi is one of the most widely spoken languages in the world but it is a complex and resource-constraint language. Thus, speech recognition and classification systems need to be developed for Hindi language to spread the technology and to explore more communication means. But due to its language complexity than other languages and lack of standard databases, it is quite challenging to develop such systems. Deep learning is extensively used in different research fields and has proven its prominence to a broader extent. In this paper, a seven-layer 1D-convolutional neural network HindiSpeech-Net has been proposed to recognise different speech samples of the Hindi language in the respective category. A large dataset of 2400 speech samples in the Hindi language is collected in ten different classes in real-world conditions which is further accompanied by signal filtering and augmentation to enhance the dataset for making a robust model and avoid overfitting. The collected dataset is divided into training, validation and test set which were evaluated in different performance parameters. The trained HindiSpeech-Net model achieved an accuracy of 92.92% on the test set. The proposed framework is computationally less expensive, works in real-time and is suitable for implementation in embedded systems.