A Comparative Analysis of LSTM and Transformer-based Automatic Speech Recognition Techniques

Automatic Speech Recognition (ASR) is a technology that leverages artificial intelligence to convert spoken language into written text. It utilizes machine learning algorithms, specifically deep learning models, to analyze audio signals and extract linguistic features. This technology has revolutionized the way that people interact with voice-enabled devices, enabling efficient and accurate transcription of human speech in various applications, including voice assistants, captioning, and transcription services. Among previous works for ASR, Long Short-Term Memory (LSTM) networks and Transformer-based methods are typical solutions towards effective ASR. In this paper, the author focuses on an in-depth exploration of the progression and comparative analysis of deep learning innovations within the ASR domain. This work starts with a foundational historical perspective, mapping the evolution from pioneering ASR systems to the current benchmarks: LSTM networks and Transformer-based models. The study meticulously evaluates these technologies, dissecting their strengths, weaknesses, and the potential they hold for future advancements in ASR.