Chunk Size Scheduling for Optimizing the Quality-Latency Trade-off in Simultaneous Speech Translation
Iqbal Pahlevi Amin, Haotian Tan, Kurniawati Azizah, Sakriani Sakti · 2024
This study addresses the quality-latency trade-off in simultaneous speech translation (SimulST) by proposing dynamic chunk size scheduling during evaluation. Traditional approaches wait for full speech before the translation begin, leading to delays. SimulST aims for minimal delay and by starting the translation incrementally. The proposed method schedules changes to chunk size during evaluation to gradually reduce predefined chunk size, aiming to maintain trans-lation quality while reducing latency. Evaluation using the MuST-C v2.0 tst-COMMON dataset shows promising results, especially in high latency regimes for English-to-German translations.