BEHIND THE SCORES

Hear it for yourself.

Five recordings. Seven speech models. The same audio, with every word of their actual responses.

Loading recordings and model responses…

01 Choose a recording

This is the exact 16 kHz evaluation audio. Each model received the same chunks of up to 20 seconds.

02 The edited reference

FULL RECORDING · CONTRIBUTOR-EDITED

These are submitted, model-assisted transcripts corrected by a contributor. They have not been independently adjudicated by native speakers. References cover the full recording; word-level and chunk-level reference timing is unavailable.

03 Compare the responses

Lower error rates are better

Choose any two models. Error rates below are for this recording; the rank beside each name is its position across all five recordings.

MODEL APinned model revision

Normalized word error rate—

Normalized character error rate—

FULL RECORDING OUTPUT

MODEL BPinned model revision

Normalized word error rate—

Normalized character error rate—

FULL RECORDING OUTPUT

Outputs are preserved from the evaluated run, without spelling fixes. Full-recording output joins nonempty chunk responses in order. WER counts substitutions, deletions, and insertions relative to the normalized reference; repeated extra words can push it above 100%. CER applies the same idea to characters. Neither is an accuracy percentage.

04 Look inside each chunk

Expand a chunk to see exactly what each selected model returned. Timestamps mark audio input boundaries, not transcript alignment. “Jump to start” moves the player; use its play control to listen.