Hume AI's Audio Decision API is a real-time audio analysis model that answers custom questions about speech and sound without a transcript. Launched in November 2026, it is built for product teams, call centers, and trust-and-safety reviewers who need structured signals from raw audio instead of a text readout. It competes with transcript-based classifiers from AssemblyAI and Deepgram by skipping the text step entirely.
Core Features
- Processes the raw waveform directly, returning probability scores for yes/no, multiple-choice, scale, and span questions with no retraining.
- Detects caller frustration, flags AI-generated voices, and classifies emotion in real time across long recordings.
- Searchable audio: index 40,000 calls and surface the relevant clips in minutes through a single query.
- Delivered as a low-latency API and SDK, so it drops into existing contact-center and moderation stacks.
- Returns calibrated confidence scores, so downstream rules can act only when the model is sure enough.
- Streams results for live calls, flagging issues as they happen rather than after the fact.
Use Cases / Best For
- Support leaders measuring customer sentiment across thousands of calls without manual review.
- Trust-and-safety teams spotting synthetic or cloned voices in uploaded media.
- Researchers who need reproducible, scored audio labels rather than fuzzy transcripts.
Pricing
Hume AI offers a free developer tier and usage-based API pricing, with enterprise plans available on request. See our AI audio tools roundup for alternatives.
Our Take
Best for teams that already sit on large audio archives and want scored answers fast. The trade-off is that 70.7% human agreement, while strong for an audio-native model, still trails careful human review on nuanced calls.
