Development of noise robust real time automatic speech recognition system for Kannada language/dialects

Abstract

Although significant progress has been made in the development of automatic speech recognition systems, performance degradation remains a challenge in uncontrolled environments. Several decades ago, signal processing -based speech enhancement techniques played a crucial role in automatic speech recognition system performance. These techniques primarily relied on the short -time Fourier transform magnitude, leaving out the short -time Fourier transform phase due to its complexity and unstructured representation. The signal processing -based speech enhancement techniques revealed difficulties in effectively combining the magnitude and unstructured phase features. In this work, we propose a technique that combines magnitude and phase features using deep neural networks to reconstruct enhanced speech for a noise -robust automatic speech recognition system. Our objective is to provide real-time access to agricultural commodity prices and weather information in Kannada language/dialects. We apply the proposed noise elimination algorithm to enhance a degraded Kannada speech database, integrating it into the real-time spoken query system before the speech feature extraction phase. Additionally, we explore the efficacy of time delay neural network and long short-term memory acoustic modeling techniques. By combining the evidence from the proposed noise reduction algorithm and time delay neural network, we achieve a relative reduction of 1.59% in word error rate compared to the earlier spoken query system. The automatic speech recognition model with the lowest word error rate is utilized in the newly developed spoken query system, which is tested with 500 farmers/speakers of Karnataka state under uncontrolled environments. To our knowledge, this represents the best -published result for a Kannada automatic speech recognition system as a real-time application serving society.