Hume launches audio decision model: capable of detecting sighs, interruptions, and irritable tones, and even providing direct scoring.

Beating AI News Flash: Hume has launched the Audio Decision API, which allows users to ask questions directly about audio recordings. For example, "How many people are speaking?" "Did the customer service agent interrupt the customer?" "Does the customer sound frustrated?" It can also pinpoint the exact times when laughter or sighing occurs. Users can ask up to 25 questions at once, and the model can deliver judgments and probabilities without needing to transcribe the text first.

Previously, determining the number of speakers, emotional shifts, interruption behavior, and background noise in a call typically required combining multiple specialized models. Each additional analysis task could require overhauling the entire pipeline. Hume hands these judgments to a single model that listens to the recording directly and answers questions. Developers only need to modify their questions to analyze new voice characteristics and conversational behaviors, without retraining models or designing specialized classifiers. A single interface can cover a wide range of audio analysis tasks.

The model was trained on over 10 million human judgment data points, attempting to predict how ordinary listeners would react to the same audio. For example, outputting 70% "frustrated" means that approximately 70% of listeners are expected to find the speaker's tone frustrated. In official preliminary tests, the model achieved 70.7% agreement with multi-rater consensus, higher than the 56.5% agreement of a single human rater.

Hume's published average score across 10 audio benchmarks is 82%, higher than Gemini 3.8 Flash's 78% and Qwen3.8-Omni-Flash's 76%.

The API currently still requires permission to access, and is priced at $0.47 per million input tokens. According to the official example, analyzing a 5-minute call costs approximately 1.2 cents. A single request supports audio from 5 seconds to 5 minutes, and real-time analysis requires continuously submitting short audio segments.