Frontier AI EvalsSign up
VOICE.TXT - Notepad

Voice and health

Sponsored by

Check that every name, dose, code and number in noisy or accented speech comes through exactly, or that an agent can finish a whole clinical admin workflow. Leaderboards report low word error rates and high subtask scores, yet one wrong digit or one missed step still breaks the record.

Benchmarks to know

  • VoiceCodeBench Vals AI

    A recording passes only if every code and identifier is exact, which exposes errors a low word error rate hides.

  • Speech to Text Artificial Analysis

    Ranks dozens of transcription models on agent talk, speeches and earnings calls, with speed and cost beside accuracy.

  • HealthAdminBench Stanford

    Agents work prior authorizations and appeals in mock health portals; a task counts only if every step is done.

  • τ-Voice Sierra

    Voice agents fail more customer-service tasks than text agents once callers bring accents, noise and interruptions.

  • MedCode Vals AI

    Coding a full hospital stay, models miss secondary diagnoses, especially mental-health and social-factor codes.

  • MLCR-AA Artificial Analysis

    Answering questions over long medical records, models leave out evidence or contradict the file.

Bring an eval in this track, or in any other, on Saturday, December 5, 2026 at Cornell University.