Hume Launches Its Audio Decision API, Answering Questions About Recordings with Probabilities
On November 2, Hume AI unveiled an API that analyzes audio directly, without transcription, and estimates the share of human listeners who would answer yes to a question (“does the caller sound annoyed?”). Access is waitlisted at $0.47 per million input tokens.
Hume AI, a developer specializing in vocal analysis, unveiled the Audio Decision API on November 2, 2026. The concept: send an audio recording along with natural-language questions (“How many people are speaking?”, “Is anyone laughing?”, “Does the caller sound annoyed?”) and receive a probability for each. The model listens to the audio directly, bypassing transcription that would otherwise strip out tone, pauses, or overlapping speech.
Numbers, Not Sentences
The documentation stresses this point: the API generates neither written answers nor explanations. For a yes/no question, it returns the estimated share of human listeners who would answer yes. A value close to 0.5 means that human listeners would be divided, not that the model is uncertain. For a multiple-choice question or a scale, it provides a probability for each option; for an event (such as laughter or background noise), it outputs time intervals, a feature that remains experimental. It is up to the application’s code to set its own thresholds and make decisions.
| Element | According to Hume’s documentation |
|---|---|
| Input | One MP3 or WAV file per request, from 5 s to 5 min |
| Questions | Multiple per request, processed in a single inference pass |
| Question types | Yes/no, choice, scale, intervals (experimental) |
| Pricing | $0.47 per million input tokens, output answers are free |
| Access | Upon request, with a one-million-token trial credit |
The announcement highlights the commercial pitch: what previously required “more than ten models” can now be handled in a single API call. Target use cases include voice agents (detecting whether the agent interrupts callers or sounds natural), call centers, and analyzing long hours of recordings.
Performance Yet to Be Confirmed
Hume claims that its model, trained on more than 10 million human judgments, agrees with a panel of raters in 70.7% of cases, compared to 56.5% for an individual human rater. It also claims the model outperforms Gemini 3.8 Flash and Qwen3.8-Omni-Flash across an average of ten benchmarks, with a median latency of 0.23 seconds. These results were published by the developer and have not been independently verified.
Our take
Returning an estimate of what human listeners would think, rather than an “angry” label presented as objective fact, is the right approach: it makes potential disagreement visible and leaves the final call to the developer. Yet it does not solve everything. Inferring emotion from vocal cues remains fragile, and such scores can easily be used to evaluate employees or customers without their knowledge. In Europe, the AI Act already bans emotion recognition in the workplace and in educational institutions, except for medical or safety reasons: companies looking to analyze their own support agents will need to tread carefully. For evaluating voice agents, however, this tool could quickly prove useful—provided Hume’s figures hold up in real-world conditions.
Sources
- Hume AI on X — Introducing the Hume Audio Decision API (November 2, 2026) Primary source
- Hume AI — Decision Model API (product page) Primary source
- Hume — Audio Decisions, documentation for coding agents Primary source
- Hume — Audio Decisions, billing Primary source
- Regulation (EU) 2024/1689 on Artificial Intelligence (AI Act), Article 5 Primary source — Prohibition of emotion recognition in the workplace and in education
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
