Measure matched speech clauses before judging voice stability

Measure the same spoken clause across independent generations when diagnosing intermittent voice deepening. A whole-response pitch median can conceal a low acknowledgment followed by a higher question, so align measurements to voiced speech and report the affected opening separately. Verify the low estimate against its octave-up alternative and count independent utterances rather than overlapping analysis frames. Interleave baseline and candidate requests so changes over the run cannot masquerade as a setting improvement. These measurements establish acoustic variability, not speaker identity or an internal decoder cause.

Applies when: synthesized speech changes tone locally within an otherwise intelligible response. Keywords: speech synthesis pitch acknowledgment clause alignment repetition interleaved acoustic measurement