Whisper Tiny · decoder evidence

Watch a transcript become readable through depth.

Three frozen ASR runs show early readability, late emergence, and tokenizer-piece behavior. Select a report, then hover or focus any layer slice for its exact rank and probability.

ASR findings

Open detailed cached explorer

Lower rank is stronger readability. The final HEAD slice is the model's actual output distribution; earlier slices are fitted readouts.

Loading recorded reports…

What this lens can—and cannot—say

Rank measures readability. It asks where the realized output sits in a fitted vocabulary-wide readout at a layer.

Rank is not calibrated confidence. A strong intermediate rank does not prove conscious consideration, causation, or correctness.

Errors remain evidence. Generated paths, reference mismatches, token boundaries, and held-out status stay visible in each report.