A multicriteria comparison of end-to-end and cascade speech-to-text translation models

Abstract

This paper presents a thorough examination of two prominent speech-to-text translation (STT) models: the end-to-end (E2E) model and the cascade model. STT is a critical technology in today’s multilingual society, facilitating communication across language barriers. The study focuses on comparing these models using a multicriteria approach to evaluate their effectiveness in translating speech to text. The E2E model represents a unified architecture that directly translates speech into text, while the cascade model involves separate modules for speech recognition and machine translation (MT). Both models have distinct advantages and challenges, which are explored in detail. Through a multicriteria comparison, this research assesses various performance metrics and criteria to determine the strengths and weaknesses of each model. The weighted sum method is employed to assign weights to evaluation criteria, providing a systematic evaluation framework. The findings have implications for researchers and developers in STT. By understanding the comparative performance of E2E and cascade models, researchers can make informed decisions regarding model selection based on criteria such as accuracy, speed, robustness, and resource requirements. This research advances the understanding of speech translation technologies and provides a foundation for future studies to refine evaluation methodologies, explore hybrid models, and enhance translation quality.