More diffusion steps did not reliably improve the voice match

Score the audio that exists

The complete indexed collection contained fifty-one converted variants. Each was independently embedded against its exact selected narrator and source waveform. Twenty-nine used four steps, eleven eight and eleven twelve.

Compare matched sentences

Across the same six Spanish utterances, mean speaker cosine was 0.8673, 0.8709 and 0.8501 at four, eight and twelve steps. The unconverted mean was 0.5727.

Ten thousand paired bootstrap resamples gave an eight-minus-four mean difference of +0.00356, interval [-0.00927, +0.01890]. Twelve-minus-four was -0.01722, interval [-0.03896, +0.00361]. Neither establishes a consistent similarity gain from extra steps on this small set.

Keep missing evidence visible

Thirty-six eight-/twelve-step outputs were absent for eighteen other phrase/reference choices. The study explicitly disabled those selections. A four-step score was never relabeled as another setting.

Four steps remained the lowest tested delay with useful resemblance. Spanish word preservation and conversion latency still pull in different directions; the higher similarity of an English style-transfer sample did not rescue its 100% ASR word error.

Evidence and publication boundary

This is an edited account of the recorded experiment. The linked archive is private; the methods, aggregate findings and limits are included above. Reference recordings, raw conversations, authentication material and workstation configuration are excluded.

• Source record 1