Hume AI Launches Audio Decision API to Analyze Speech and Sound
Developers ask plain-language questions about raw audio and get back probabilities, priced at $0.47 per million input tokens.
Published
Last updated / Fact-checked on
Hume AI on November 2 launched its Audio Decision API, which takes raw audio plus developer-written questions and returns probabilities for the possible answers. Access is through a waitlist, and pricing is $0.47 per million input tokens, according to Hume's launch post and billing docs.
Photo: Pexels / Pixabay
Audio Decision API at a glance
| Company | Hume AI |
| Input | One MP3 or WAV file, 5 seconds to 5 minutes |
| Output | Probabilities for developer-defined answers |
| Question types | Yes/no, choice, scale, span |
| Reported accuracy | 70.7% agreement with panel average vs 56.5% individual-rater baseline (Hume) |
| Price | $0.47 per 1M input tokens, no commitment |
| Access | Waitlist; playground available |
| Sources | Hume blog, billing and developer docs |
Hume describes the API as "the first audio-native model that answers questions about speech and sound." It presents the product as a way for developers to extract structured judgments directly from recordings, including details that might be lost when speech is converted into text. The "first" label is Hume's claim.
Instead of sending a transcript to a language model, developers submit audio and questions written in ordinary sentences, accompanied by the possible answers. According to Hume, the service returns probabilities for those options, giving applications values they can use to make decisions rather than a generated explanation.
Four Ways to Ask About Audio
Hume's launch post describes four question formats: yes/no, choice, scale and span. A yes/no question might ask whether a caller sounds frustrated. A choice question can compare descriptions of an agent's tone, while a scale can rate perceived patience. Span questions identify start and end times for events such as laughter, sighing or gasping.
The company claims the system is fast enough to answer dozens of questions every few seconds during a live call and make an archive of 40,000 recordings searchable within minutes. Those are company claims.
Hume's developer guide says that responses are numeric judgments, not transcripts or conversational replies. It says a single request contains one audio file, and questions are processed together. For live workflows, developers can send successive clips rather than a continuous audio stream.
Research and Early Evaluation
Hume ties the new model to its research on how people perceive emotion and expression in speech. In its announcement, the company says tens of millions of listener judgments informed its Real-World VoiceEQ work and helped train the Audio Decision model.
In an initial evaluation reported by Hume, the model reached 70.7% agreement with the average judgment of a human panel. The company compared that figure with a 56.5% individual-rater baseline. The figures come from Hume's own evaluation and have not been independently verified.
The comparison also needs careful interpretation. Hume explains that assessments such as whether speech sounds natural, empathetic or frustrated can vary between listeners. Agreement with a panel therefore measures alignment with those averaged perceptions, not proof that every output correctly identifies a speaker's emotional state.
Access, Limits and Pricing
Developers can request access through a waitlist, according to Hume's launch post. The company also offers a playground for trying questions against real-time audio and uploaded recordings. It describes live and ambient applications as a longer-term opportunity as capabilities develop.
The developer guide specifies one MP3 or WAV file per request, lasting between five seconds and five minutes. Shorter or longer audio is rejected. The guide also advises developers to treat span questions as experimental, most reliable for vocal bursts and sound events.
Hume's billing documentation sets pricing at $0.47 per million input tokens without a commitment. Hume says tokens are calculated from audio duration plus question and context text; returned answers are not billed. Every second of audio counts, including silence.
Once access is granted, an organization receives a trial credit of one million input tokens, according to Hume's billing guide.
What to watch
- General availability. When access opens beyond the waitlist.
- Independent tests. Whether outside evaluations reproduce the 70.7% agreement figure.
- Live use. How the API performs on real-time calls, which Hume calls a longer-term opportunity.
- Span reliability. Whether span questions move past experimental status.
More coverage is on our Artificial Intelligence hub.
How we reported this
We checked the article against Hume's launch post and its developer and billing documentation. Accuracy and speed claims are Hume's own and attributed as such. We did not test the API.
Published and fact-checked November 2, 2026.
Sources
- Hume AI, "Introducing Hume's Audio Decision API: A decision layer for speech and sound," Nov. 2, 2026: https://www.hume.ai/blog/introducing-hume-audio-decision-api
- Hume developer documentation, Billing, accessed Nov. 2, 2026: https://dev.hume.ai/audio-decisions/docs/billing
- Hume developer documentation, For coding agents, accessed Nov. 2, 2026: https://dev.hume.ai/audio-decisions/docs/for-coding-agents
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.