ABOUT THE CALL
DO NOT REWIND turns a large voice-resemblance archive into ten quick “same voice or different voice?” rounds. A is always the known reference voice; B is the clip being compared.
PLAY A QUICK CASE →A stays the known reference. B rotates among accepted matches, archive candidates, other voices, and occasional Grandma bonus comparisons. The compact result stays between both clips, and the next round starts automatically after ten seconds.
The thumbnails identify the Kitboga videos being compared, and the public media index preserves the source-video URL for every excerpt. The short FLAC excerpts and video thumbnails are packaged with this static exhibit. Each excerpt retains its source-video URL and speaker label in the dataset.
OPEN THE MEDIA INDEX →Our label shows the current manual classification: same, different, or maybe. Player votes shows the running listener split saved on this device. A public version can combine visitors’ votes through a small poll endpoint.
The Grandma voice acts as a recognizable recurring-character benchmark and stops the target-speaker cases from becoming monotonous. It also gives the game a much-needed dose of chaos.
The accompanying thesis explains the speaker-resemblance evidence, the multi-model comparisons, and the limits of the analysis.
OPEN THE THESIS (PDF) →The advertised gift card has, regrettably and completely unexpectedly, already been redeemed. No refund, replacement, or accidental overpayment is available.
