Distinguish deterministic speech from stable voice quality

Removing randomness from a reference-conditioned speech generator can freeze an undesirable voice register rather than stabilize the intended voice. Verify that the effective sampling flag changes, then measure repeated native audio, phrase-dependent pitch, and natural termination separately. Identical bytes prove repeatability, not voice quality. If lower-pitch output persists with zero sampler invocations, random sampling is not necessary for that failure; investigate conditioning and model behavior instead of declaring the settings correction a voice fix. This inference is bounded to the tested phrases and runtime, and pitch alone does not establish speaker identity.

Applies when: diagnosing register changes in a generative speech server after correcting an ignored sampling override. Keywords: speech voice pitch greedy sampling deterministic conditioning repetition termination