MULTILINGUAL CODEC AUDIT · 2026
Does compression
change the words?
A controlled 18-language study of transcript stability after a full-codebook Descript Audio Codec roundtrip.
DAC9CODEBOOKS
01 / LANGUAGE VIEW
Stability across scripts
Error bars are deterministic 95% bootstrap intervals. Japanese uses a mixed script-aware token metric; grapheme error is comparable across all languages.
02 / SIGNAL ↔ TEXT
Where acoustics meet recognition
Each point is one clip. Select an acoustic measure to inspect how codec distortion relates to transcript disagreement.
03 / DISAGREEMENT LAB
Hear what changed.
See exactly where.
Filter and rank all 9,000 clips, inspect aligned insertions, deletions, and substitutions, then audition synchronized original and DAC waveforms.
Open the detailed explorer ↗04 / INTERPRETATION
Pseudo-reference,
not human truth.
The same pinned Qwen3-ASR-0.6B model transcribed originals and DAC reconstructions. The original transcript is the pseudo-reference, so these rates isolate model output instability introduced by the codec. They do not measure ASR accuracy against human labels.
Text is normalized with Unicode NFKC, case folding, control removal, punctuation-to-space conversion, and whitespace collapse. Results pool substitutions, deletions, and insertions rather than averaging percentages.
pWER = (S + D + I) / N