A Practical Evaluation Method for Long-Form Simultaneous Speech-to-Speech Translation

In three linesPractical evaluation method for long-form simultaneous speech-to-speech translation (SimulS2ST) on continuous input. Uses ASR, forced alignment, and sentence embeddings to recover timestamps and align target text to source sentences, then computes sentence-level latency and quality metrics (YAAL, xCOMET). Reveals substantial latency accumulation in current systems on long speech.

Summary generated by Claude — human-verified