Hume’s new service answers developer-defined questions directly from sound
Hume AI has announced Audio Decision API, a service that assigns probabilities to answers chosen by developers after examining raw audio. Questions can cover speakers, background sounds and perceived expression without relying only on a transcript. The same recording can answer multiple questions, with applications ranging from call-archive searches to comparisons of generated voices.
Artificial Intelligence··Morning
Developers choose questions and possible answers
Hume AI, which develops artificial intelligence for speech and sound, has announced Audio Decision API. This programming interface examines raw audio and returns probabilities for answers specified by a developer. It is designed to handle several structured questions about the same recording without relying only on a transcript. The developer supplies both a question in ordinary language and its possible answers, rather than asking for an open-ended description of everything in the clip.[1], [2]
Questions range from speaker counts to moments of laughter
Supported answers include yes or no, a choice from a list, a rating scale or a time span within a recording. Hume gives examples such as identifying speakers, detecting background noise and locating laughter, sighs or gasps. Questions can change without training a separate classifier for each new task. Developers can add context and connect the returned probabilities to business rules using thresholds. A customer-service application can combine those audio judgments with a transcript as input to a language model.[1]
Hume separates sound detection from listener perception
Hume distinguishes relatively objective questions, such as whether a siren is present, from judgments about how empathetic, natural or frustrated a voice sounds. Listeners can disagree on those perceptions. The company recommends checking evaluations against representative human ratings. Its proposed uses include comparing generated voices, finding calls in an archive and following expression across successive audio segments. A playground supports real-time audio and uploads, while programming-interface access requires joining a waitlist. Broader live and ambient applications remain longer-term plans.[1]
Related columns
For more information on this topic, you can read the related columns.
