Hume AI opens API that answers questions about voice recordings, matching humans 70.7% of the time

Anyone who has built voice agents or call analytics knows the problem: a transcript loses half the meaning. Text will not reveal that a customer sighed, raised their voice in irritation, or that a child was crying in the background. The usual remedy is a collection of separate classifiers for emotion, noise, and speaker diarization, each requiring its own labels and fine-tuning. Hume AI proposes replacing this entire menagerie with a single model.

What exactly was released

The product is called Audio Decisions. It takes a recording, your question about it, and a list of allowed answers. It returns probabilities for those answers and a confidence score. Four answer formats are supported: yes or no, multiple choice, a rating scale, and a time span within the recording—for example, when a person laughs or sighs.

You can ask about language and speech, the number and changes of speakers, emotions, the acoustic environment, non-speech events, conversation dynamics, customer intent and call outcome, as well as social cues. The question set can be changed on demand without additional labeling or fine-tuning. The confidence threshold can be tied to the customer's rules: for example, route a call to a human only if the model is at least 80% confident that the caller is irritated.

What this looks like in practice

Hume offers a retail scenario. A voice agent talks to a customer while simultaneously asking the audio model questions: whether the customer sounds uncertain, whether irritation is audible, how many people are speaking, whether there is background noise, and what the customer wants. These answers, together with the transcript, are passed to a language model, which then decides whether to wait, ask a clarifying question, or hand the conversation over to a human agent.

The architecture makes sense. The audio model acts as the “ears,” producing structured output, while the LLM acts as the brain that makes the decision. There is no need to teach the language model to listen; it only needs clear signals as input.

According to the company, during a live call the model answers dozens of questions every few seconds, while an archive of 40,000 recordings becomes searchable within minutes. For contact-center quality assurance, this creates an opportunity to move from reviewing a small sample of calls to reviewing all of them.

How good is it?

Hume compares the model with Gemini 3.8 Flash, Qwen3.8-Omni-Flash, Inkling-Small, gpt-audio-1.5, Gemini 3.5 Flash Lite, and Nemotron 3 Nano Omni, placing its own model ahead of all of them. In a comparison with people, the model agreed with the average judgment of a group of listeners in 70.7% of cases, while an individual evaluator did so in 56.5%. In other words, the model is closer to the consensus than a randomly selected person.

That is a strong argument, but it comes with caveats. The company published the benchmarks itself, and the announcement does not make the question set or dataset composition clear. Moreover, 70.7% is nowhere near 99%. Emotions and intentions are subjective, and even people disagree about them. For tasks such as legally consequential employee evaluations, this level of accuracy is a reason to keep a human in the loop.

What developers and product teams should do

Access is currently available only by joining a waitlist, and the price is 47 cents per million input tokens. Audio is tokenized densely and call volumes can be high, so estimate the cost using your own volumes in advance.

A practical plan:

• if you operate a contact center or voice agent, collect 100–200 of your own calls with expert assessments and test the model on them rather than relying on the demo;
• compare the results with Gemini or gpt-audio using the same questions: multimodal models can also answer such questions through prompting, though without the claimed confidence calibration;
• define thresholds in advance: decide which cases go to a human and which are handled automatically.

If the reported figures hold up on your data, Audio Decisions solves the most unpleasant part of the project: labeling and maintaining a collection of narrow classifiers. If they do not, you will lose a couple of evenings to testing but retain a realistic understanding of the true ceiling of voice analytics.