Second Ear

Second Ear: an AI check interpreter for immigration hearings In US immigration court, the record is the interpreter’s English. If “three times” becomes “two times,” the transcript keeps “two times,” and appeals fail without specific mistranslations (Deh v. Blanche, 6th Cir. 2026). Second Ear reviews the hearing audio and returns ranked, checkable discrepancies: timestamp, words spoken, words interpreted, a proposed correction, and both audio clips, so a qualified interpreter can confirm each in minutes. Meta models do the work: Muse Voice Transcribe hears both languages; Muse Spark 1.3 runs 7 error-family judges, a holistic pass, and a 3-vote verifier; Omnilingual ASR re-transcribes each flag on a local GPU; Muse Spark 1.2 listens to both clips to separate interpreter errors from transcription errors; and Muse Wonder reads the whole hearing to find inconsistencies the interpretation created and link errors to the judge’s decision. On a hash-locked test of 6.7 hours with human labels, Second Ear reached recall 0.77, against 0.003 for a one-prompt Muse Wonder baseline. On 30 mock hearings it reached recall 0.95 and F1 0.84. It linked 37 of 38 errors a decision relied on. We report where it falls short: on text alone it matches one good prompt, and on real noisy speech it makes 88 false alarms per hour. Live mode supports consented prep sessions only.