Penny speech comparison
Compare the same Penny and Sparkle lines, including the story line that stopped in the audit. Voice, speed and wording stay the same. This test uses no camera or microphone.
Each pass reads 12 short lines per selected profile and makes paid Azure requests. Start with one pass; run three for a steadier comparison. Profiles alternate order. Rate the voice and pronunciation after listening.
Ready when you are.
| Profile / line | First sound | Result | Words / voice / name |
|---|
Listening and interaction check
This output test cannot tell us whether a short child answer was recognised. Test listening separately, using the same microphone, distance and speech clips with PENNY_TRANSCRIBE_MODEL=whisper-1, then gpt-4o-mini-transcribe. Restart between these two runs. Keep speech profile A for both.
At the matching app prompts, say: “Krish”, “Four”, “Blue”, “one, two, three, three fireflies”, “stars”, “Can we add the moon?”, “another story”, and “back to games”. Include a quiet background-noise interval. Compare the audit's child_transcript with the accepted app action. Correct text with no action is a controller failure; missing or wrong text is a listening failure. Noise must never advance a prompt.
See the speech test guide for the full procedure and pass criteria.