Size and Inference Time Optimized Automatic Speech Recognition Model

In this work, development of an efficient optimized Automatic Speech Recognition (ASR) system is done by leveraging methodologies. This study focuses on knowledge distillation, a technique where a master model’s knowledge is transferred to a smaller slave model. This process aims to reduce the computational complexity and memory footprint of the ASR system, enabling efficient deployment on resource-constrained devices. To further optimize the slave model, pruning, quantization, and clustering techniques are employed. Pruning removes less important weights, quantization reduces weight precision, and clustering groups similar weights. These approaches collectively contribute to minimizing the size of the slave model and improving its inference time. Additionally, the study investigates various feature extraction techniques, including spectrogram, mel spectrogram, and MFCC, to represent audio data effectively. Data augmentation techniques, such as time stretching, pitch shifting, noise injection and reverberation, are employed to enhance the training process and improve the system's robustness. The study’s outcomes are evaluated using standard performance metrics, such as word error rate (WER) and accuracy, to measure the effectiveness of the optimized ASR system. Experimental results demonstrate the potential of the proposed methodologies in achieving accurate speech recognition while addressing the challenges of model size and inference efficiency. The developed ASR system holds promise for real-world applications, where resource constraints are a significant consideration.