Hume announced the Audio Decision API on X, saying developers can ask questions such as how many speakers are on a recording or whether a caller is frustrated, that what used to take 10+ models is now one call, and that the model was trained on 10M+ human judgments; access is by request.
Hume, the voice AI company, launched its Audio Decision API on Monday, 2 November, and said in its launch post that in an initial evaluation the model achieved 70.7% agreement with a human panel’s average judgment, compared with 56.5% for the individual-rater baseline.
That is a gap of 14.2 percentage points, by our arithmetic. The figures are Hume’s own and have not been independently tested.
The key insight: Hume launched an Audio Decision API that returns probabilities for answers to questions developers ask about speech and sound, and says that in an initial evaluation the model agreed with a human panel’s average judgment 70.7% of the time, against 56.5% for a single rater.
What the API Does
Hume describes the product as “the first audio-native model that answers questions about speech and sound”. Developers send raw audio, write a question as a sentence and list the possible answers; the API returns probabilities over those answers, according to the launch post.
Questions can be yes or no, multiple choice, rating scales or spans across a clip, the company says, and teams can ask several questions at once and change them in real time. Its examples range from whether a voice is AI generated to whether a caller is frustrated or a siren is audible in the background.
In its X announcement, Hume said the model was “Trained on 10M+ human judgments” and that “What used to take 10+ models is now one call”. The launch post says tens of millions of listener judgments informed Hume’s Real-World VoiceEQ and helped train the model.
Agreement with a human panel’s average judgment in Hume’s initial evaluation: 70.7% for the Audio Decision model and 56.5% for the individual-rater baseline (Hume, 2 November 2026). The 14.2-point gap is our arithmetic.
Speed and Scale Claims
Hume says the system is fast enough to ask dozens of questions every few seconds on a live call, and fast enough to make an archive of 40,000 audio recordings searchable in minutes. It also says the model outperforms existing audio methods on accuracy, speed and cost, without publishing that benchmark in the post.
Access is through a waitlist, the company says, with a playground for realtime and uploaded files. The launch post does not give pricing.
Why It Matters for Voice Agents
Hume’s pitch is that transcripts drop the pause, the sigh and people talking over each other. The post says a cascaded voice agent could weigh an audio judgment alongside the transcript and its business rules before it responds, and that contact centers could search an entire call library for, in its example, every call where the customer sounded frustrated.
In its section on evaluating voice models, Hume adds a caution: “Teams should validate their evaluations against representative human ratings.”
As we read it, Hume is positioning the API as a judgment layer that sits beside transcription and voice models: the launch post says an agent could weigh an audio judgment alongside the transcript and its business rules.
As we read the evaluation, the comparison point Hume reports is a single human rater rather than another model; the launch post does not compare the model with other audio systems on that metric.
Hume, launch post, on evaluating voice models, 2 November 2026
“Teams should validate their evaluations against representative human ratings.”
Three Implications
FOR CONTACT CENTERS Hume says teams can run questions across entire audio libraries and find, in its example, every call where the customer sounded frustrated.
FOR VOICE-AGENT BUILDERS The launch post says a cascaded voice agent can use the API’s audio judgments alongside the transcript and its business rules before responding.
FOR BUYERS Access is through a waitlist and the launch post gives no pricing, so the cost claims Hume makes cannot yet be checked.
The framework reads: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, Hume’s API is pitched as a component of that harness for voice agents: a decision layer that turns audio into structured answers an agent can act on. The framework is the owner’s argument and does not discuss Hume.
Hume says its model was trained on human listener judgments; this Business Pill explains why data quality sets what a model can learn.
What Is Not Established
Hume does not say how many recordings or raters the initial evaluation used, which languages it covered, or what the API costs. The 10M+ figure is from Hume’s X post and the 70.7% and 56.5% figures from its launch post; neither has been independently verified. The 14.2-point gap is our arithmetic. We did not contact Hume.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Hume launched an Audio Decision API that answers developers’ questions about recorded or live speech with probabilities, and says its model agreed with a human panel’s average judgment 70.7% of the time against 56.5% for a single rater.
A note on sourcing. Product details, evaluation figures and quotes are from Hume’s 2 November launch post and its X announcement. The 14.2-point gap is our arithmetic. We did not contact Hume.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
