Validate voice conditioning across wording as well as seeds

Validate voice-conditioning changes across wording as well as seeds.

When a speech model changes register, repeat the affected wording under matched random seeds, then extend validation to varied wording. Removing reference-speech continuation while retaining the same speaker embedding can reduce a specific low-pitch pattern, yet other phrases can still produce low onsets or delayed termination. A perfect result on one sentence therefore establishes a useful intervention, not universal voice stability. Measure phrase-level pitch, natural end tokens, and quiet tails separately; pitch alone does not establish speaker identity. Preserve native audio and effective prompt metadata so improvements cannot be confused with a changed reference voice.

Applies when: evaluating a reference-conditioned speech generator for within-call voice consistency. Keywords: speech voice conditioning sampling seeds wording pitch termination quiet tails