Hume Audio Decision API

Hume Audio Decision API (the docs also call it the Audio Decisions API) is an audio-native decision model launched by emotion-aware voice AI company Hume AI on November 2, 2026. Developers send raw audio plus plain-language questions ("Does the caller sound frustrated?", "Is this voice AI-generated?", "Is there a siren in the background?"), and the model returns probabilities over the answers instead of a transcript or a chat reply.

Hume calls it "the first audio-native model that answers questions about speech and sound," aimed at the gap where voice agents only read transcripts and miss tone, pauses and sighs.

How it works

• One request = one audio file + many questions, all answered in a single inference.
• Four question primitives:
  • noul: yes/no, returning the share of human raters who would say yes;
  • choice: pick among unordered options, with probabilities;
  • score: position on an ordered scale, with score, confidence and distribution;
  • span (experimental): time ranges where an event occurs, best for vocal bursts and sound events.
• Numbers only: no transcript, explanation or reply. Your code thresholds, ranks or combines the outputs, which makes it easy to wire into business rules or agents.
• An optional context string describing the recording can improve results.

Specs and pricing

Reported results: trained on 10M+ human judgments, it agrees with human raters about 70.7% of the time, versus 56.5% agreement between individual listeners.

Typical use cases

• Call-center QA: flag frustrated customers, repeated requests for a human, or agents talking over callers across large call archives.
• Voice-agent evaluation and live monitoring: query clips every few seconds and trigger escalation or strategy changes.
• Trust and safety: detect AI-generated voices or unusual background sounds.
• Audio archive search: tag tens of thousands of recordings by tone and events so they become searchable.

Getting started

1. Request access on the Hume Platform and get an API key (1M-token trial credit).
2. Read the Audio Decisions docs at dev.hume.ai and estimate cost with the input_tokens endpoint.
3. Send one MP3/WAV with a few noul/score questions via cURL or the SDK and inspect the probabilities.
4. Set thresholds for your use case and route results into tickets, alerts or agent decision logic.