Benchmark latency when swapping voice references

A voice-reference swap is also a performance experiment. In reference-conditioned speech synthesis, changing only the prompt recording can change output speaking rate and code count, so identical model and decoder settings need not produce identical total generation time. Hold phrases, repeats, model, and machine fixed; report first-chunk and full-generation medians separately, with output duration and intelligibility. A more varied voice can still be viable if its measured cost fits the declared latency ceiling. A small isolated run does not predict production tail latency or prove perceptual naturalness.

Applies when: choosing a voice reference for a latency-sensitive speech service. Keywords: TTS voice reference latency matched phrases first chunk generation duration prosody