Hume unveils an audio decision model that can detect sighs, interruptions, and frustrated tones, and also directly score audio inputs.

Beating AI Express (Dongcha) reports: Hume has launched the Audio Decision API, which enables direct queries of audio recordings. Supported questions include "How many people are speaking?", "Did the customer service agent interrupt the client?", and "Does the client sound agitated?" The tool can also pinpoint the exact timestamps of laughter and sighs. Users can submit up to 25 queries at a time; the model delivers judgment results and probabilities without prior transcription of audio content. Previously, analyzing metrics like number of speakers, emotional shifts, interruptions, and background noise in a call required combining multiple specialized models, with each new analysis task often needing adjustments to the entire workflow. Hume centralizes these judgments into a single model that processes audio directly to answer queries. Developers only need to modify their questions to analyze new audio features and conversational behaviors, eliminating the need to retrain models or design custom classifiers. A single API can cover a wide range of distinct audio analysis tasks. The model is trained on over 10 million human judgment data points, aiming to predict how average people would react to the same audio. For example, a 70% "agitated" output indicates approximately 70% of listeners would perceive the speaker’s tone as agitated. In official preliminary tests, the model’s agreement rate with multi-reviewer consensus reached 70.7%, outperforming the 56.5% average of individual human raters. Hume reported an average score of 82% across 10 audio benchmarks, higher than Gemini 3.8 Flash’s 78% and Qwen 3.8-Omni-Flash’s 76%. The API is currently available via permission application only, priced at $0.47 per million input tokens. Per official examples, analyzing a 5-minute call costs roughly $0.012. It supports audio clips from 5 seconds to 5 minutes per request; real-time analysis requires submitting short audio segments sequentially.