Percentile-Based Evaluation for Speech-to-Speech Agent Prosody and Rhythm

This method evaluates spoken dialogue systems by comparing S2S output waveforms against matched human reference regimes using 4,000 hours of conversation data. The protocol uses percentile deviations in F0, speech rate, and pause duration to detect out-of-regime rhythmic or expressive errors.