Hume launched its Audio Decision API for speech and sound analysis. It answers custom questions about raw audio in real time. The API returns probabilities for user-defined answers without retraining. It processes live calls fast enough for dozens of questions every few seconds. It also searches large audio libraries in minutes. The model matches human judgment on emotional expression at 70.7% agreement. It outperforms other audio models on accuracy, speed and cost. Developers define questions as sentences with possible answers. It supports yes/no, multiple choice, rating scales and time spans. Use cases include evaluating voice models and building perceptive agents. Access requires joining a waitlist.
What changed
Audio analysis required separate pipelines for each question, with manual labeling and weeks of setup.
What it unlocks
Asking custom questions about raw audio in real time and getting structured answers with confidence scores.
- 70.7% agreement with human panel
- 56.5% individual-rater baseline
- 0.82 benchmark score at 0.23s latency
- 40,000 audio recordings searchable in minutes
What you need to act on it
- waitlist access
- API key
Sources