VOICE A/B LISTENING SCORECARD
Proposed evaluation template | Useful in Practice

Define before testing:
Task and primary outcome:
Voice model versions and settings:
Fixed scripts and reference pronunciations:
Anonymous clip labels and counterbalanced order:
Playback equipment and conditions:

Repeat this record per script and listener:
Listener ID (anonymous):
Script ID:
Clip order:
Clip labels (keep voice identity hidden):
Pronunciation accuracy (1=frequent errors; 5=correct throughout):
Clarity (1=difficult to understand; 5=clear throughout):
Task comprehension (record the final time/number/next action):
Preference (first/second/tie) and reason:

Capture separately per request:
Voice model, region, settings, audio format, timestamp:
Time to first byte (ms):
Total request time (ms):
End-to-end playback delay, if measured (ms):
Request errors or retries:

Summarize completed comparisons, missing data and disagreements.
A listening pilot does not prove production conversion lift.