Analyzing Multilingual Automatic Speech Recognition Systems Performance

Understanding spoken language, or transcribing the spoken words into text, was one of the earliest goals of computer language processing and falls under the realm of speech processing. Speech processing in itself predates the computer by many decades. Speech being the most important and most common means of communication for most people, is always in need of necessary technology advances. Therefore, in the recent decades there has been great interest in techniques including automatic speech recognition (ASR), text to speech etc. This research is focused around English and the scope needs to be expanded to other languages as well. In this study we explore several open-source ASR systems that offer multilingual (English and Spanish) models. We discuss various models these ASR systems offer, evaluate their performance. Based on our manual observations and using automatic evaluation metrics (the word error rate) we find that Whisper models perform the best for both English and Spanish. In addition, it supports a multilingual model that has the ability to process audio that consists of words from both English and Spanish.