Hume introduces an API for asking questions about sound, with access by waitlist

Hume introduced its Audio Decision API on November 2, giving developers a way to ask questions about raw audio and receive probabilities for answers they define. The company opened a waitlist for access, pitching the tool for decisions that transcripts alone can miss, from overlapping speakers to a caller who sounds frustrated.

API access requires joining the waitlist, while a playground offers real-time and file-upload experiences with custom questions. Hume presents the product as a way to turn speech and sound into structured judgments that software can use directly. Rather than asking a model to write an explanation, a developer supplies the question and the allowed answers.

The question becomes the configuration

Hume says developers can ask several questions about the same recording, change them in real time and add context to each request. Adding a question means adding a line of JSON, a structured data format, rather than labeling examples or retraining a model. The interface supports four kinds of response, covering both judgments about a clip and locations within it:

• Yes or no: assess whether a customer sounds frustrated.
• Multiple choice: select the description of an agent’s tone from a supplied list.
• Rating scale: score how patient an agent sounds using the developer’s scale.
• Time spans: return start and end times for laughter, sighs or gasps.

The same recording can be queried for language, accent, background noise and emotional expression. Developers can set confidence thresholds that connect the responses to business rules. Hume’s design sits between fixed classifiers, which handle predefined tasks, and generative models, which produce open-ended text even when an application only needs a structured answer.

Alongside the transcript, not instead of it

Hume’s retail customer-service example shows where those answers would go. A voice agent could check whether someone sounds uncertain or frustrated, count speakers, assess background noise and identify whether the caller wants a return or cancellation. A language model could then receive both the transcript and those audio judgments, using business rules to wait, clarify or escalate.

The proposed benefit is access to cues that words alone leave out. Hume gives two examples: someone saying they are fine while sounding frustrated, and a pause that signals thinking rather than the end of a turn. Despite the real-time playground, the company describes live and ambient applications as a longer-term opportunity as real-time capabilities develop.

For recorded audio, Hume proposes bulk searches for calls with interruptions or frustrated customers, plus tracking expression across successive segments. Voice-model teams could also compare whether generated speech sounds natural, conveys the intended emotion or fits a situation. Those judgments could guide training feedback and data selection; Hume advises teams to validate evaluations against representative human ratings.

70.7%Audio Decision Model

Hume reports 70.7% agreement with a human panel’s average judgment.

56.5%Individual-rater baseline

The individual-rater baseline agreed with the panel’s average judgment 56.5% of the time, according to Hume.

A listener’s judgment is the target

Hume distinguishes relatively objective tasks, such as counting speakers or detecting a siren, from perceptions such as frustration, empathy or naturalness. Listeners do not always agree on the latter. The company says tens of millions of human judgments informed its model, drawing on panels that rated emotional expression, vocal delivery, naturalness and conversational fit.

Separately, Hume’s product page reports an average score of 0.82 across 10 benchmarks and median latency of 0.23 seconds per screen. Those are company-reported performance figures. The listener-agreement result has a more specific meaning: it measures agreement with a panel’s average judgment, rather than providing one accuracy score for every question developers might ask.