Track and Noise Separation Based on the Universal Codebook and Enhanced Speech Recognition Using Hybrid Deep Learning Method

• Published in: IEEE access Vol. 11; pp. 120707 - 120720
• Main Authors: Kumer, S. V. Aswin, Gogu, Lakshmi Bharath, Mohan, E., Maloji, Suman, Natarajan, Balaji, Sambasivam, G., Tyagi, Vaibhav Bhushan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Blind source separation; Blind source separation (BSS) method; Codes; Deep learning; deep learning method; Generative adversarial networks; generative adversarial networks (GAN); Machine learning; multi-channel method; Noise measurement; noise separation; Speech coding; Speech enhancement; Speech processing; Speech recognition; Task analysis; track separation; voice assistant; Voice recognition
• DOI: 10.1109/ACCESS.2023.3328208
• ISSN: 2169-3536

The concept of Deep learning is a part of machine learning which is very useful nowadays to achieve accurate voice and speech recognition based on the training data by creating robust algorithms. It is also possible to separate the noise from original speech as well as the separation of tracks in particular audio signal with the help of machine learning algorithms. In this paper, the implementation is applicable for voice assistant to separate the tracks and the noises from the multiple original audio which reproduces simultaneously using the speech enhancement and universal code book. For that, the Hybrid Deep Learning Algorithm has been developed and the training data sets are also created and achieve the accuracy in the speech recognition for the variety of voice assistants. Most of the time, the voice assistant recognizes the voice with noises and musical audio which results in the malfunction of devices which can be controlled by the same voice assistant. The Generative adversarial networks from Deep learning and the blind source separation method from multi-channel model are combined to form this proposed hybrid deep learning model.