Quantitative Analysis of Error Propagation in Hindi-English Cascaded Speech-to-Speech Translation Models

Speech-to-Speech Translation (S2ST) plays a crucial role in reducing language barriers and enabling seamless communication between people from different linguistic backgrounds. Traditional S2ST systems typically rely on a cascaded design, where components like Automatic Speech Recognition (ASR), Machine Translation (MT), and Text-to-Speech (TTS) work together in sequence to complete the translation process. Cascaded models are widely used S2ST models and are prone to error propagation (EP) across the pipeline, significantly impacting translation quality. EP is discussed in various literature. However, a comprehensive quantitative analysis is not available, particularly for low-resource languages like Hindi and English. This work presents a detailed quantitative study of EP in Hindi-English cascaded S2ST models, bridging this critical research gap. This study utilizes both text-based and textless evaluation metrics such as BLEU, Translation Edit Rate (TER), and BLASER score for translation accuracy to quantify the impact of EP at various stages of the pipeline on the FLEURES dataset. The result analysis shows that due to EP, the translation quality of the S2ST model decreases in BLEU score of 11.55 for English→Hindi and BLEU score of 12.16 for Hindi→English. Similarly, reference-based BLASER decreases by 0.61 and 0.45 for English→Hindi and Hindi→English, respectively.