Robust and efficient keyword spotting using a bidirectional attention LSTM

• Published in: International journal of speech technology 2023-12, Vol.26 (4), p.919-931
• Main Authors: Swain, Om Prakash, Hemanth, H., Saran, Puneet, Kothandaraman, Mohanaprasad, Ravi, Logesh, Sailor, Hardik, Rajesh, K. S.
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Attention; Autonomous cars; Denoising; Engineering; Human-computer interface; Imperative sentences; Keywords; Machine learning; Mass media; Model accuracy; Noise reduction; Signal,Image and Speech Processing; Social Sciences; Speech recognition; Training; Voice recognition
• ISSN: 1381-2416; 1572-8110
• DOI: 10.1007/s10772-023-10067-4

Speech recognition and voice assistants have become integral in various advanced electronic devices and human-computer interfaces ranging from smartphones to self-driving cars. Personal assistants like Google Assistant and Amazon Alexa are state-of-the-art examples of personal voice assistants triggered by simple wake words. Much research has been conducted on improving wake word detection and keyword spotting techniques. Along those lines, this research aimed to build an enhanced keyword-spotting model with high accuracy and reduced training time. The proposed model involves a Convolutional Neural Network and a bidirectional attention LSTM network. Another critical area of focus in this paper is audio pre-processing, where the Spectral gating technique has been introduced for denoising the data. The aforementioned CRNN model exploiting the abilities of CNN and LSTM was trained on this denoised data. The model has trained on 35 keywords from the Google Speech Commands dataset and can identify each. The model training was performed using Google Colab. The model proposed in this paper achieved a training accuracy of 93.44% and a test accuracy of 92.94%.