Comparative Analysis for Speech Recognition Using DeepSpeech

The presented work aims to implement speech recognition systems for Hindi and English using advanced deep learning. Enhancing ASR technologies is crucial for improving digital communication's accessibility and efficiency across different linguistic contexts, which is pivotal for fostering broader adoption of technology in multilingual environments. The experiments were conducted using speech corpora such as LJSpeech(English Language) and OpenSLR(Hindi Language). The novelty of this research lies in the tailored adaptation of hybrid deep learning architectures, combining the strengths of Convolutional Neural Networks (CNNs) for feature extraction and Recurrent Neural Networks (RNNs) for sequence learning, specifically configured for the acoustic properties of English and Hindi. These architectures are integrated using connectionist temporal classification to enhance transcription precision without the need for linguistic pre-segmentation. The research findings provide evidence that the implemented hybrid deep learning architectures significantly enhance the performance of ASR systems. By specifically addressing the unique challenges posed by Hindi languages, the models demonstrate improved accuracy in transcription, particularly in environments with varied acoustic properties. The system obtained a minimum WER of 21.89. This approach showcases how advanced neural network configurations can adapt to complex linguistic features, thus reducing the Word Error Rate and improving the overall reliability of speech recognition systems in real-world applications.