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 ph…