Hume AI Launches Audio Decision API for Voice AI Queries

Hume AI introduced the Hume Audio Decision API, a single model that answers questions directly from audio, replacing what previously required more than 10 models. Trained on over 10 million human judgments, the API handles realtime or long-recording analysis and averages 0.82 accuracy across 10 benchmarks at 0.23 seconds median latency.
Event version 1.
Key insights
Single-call audio questioning
The Hume Audio Decision API answers questions such as speaker count, laughter detection, or caller frustration directly from audio. Hume states that tasks previously requiring more than 10 models now take one call.
Trained on human judgments
The model is trained on more than 10 million human judgments. It agrees with a panel of human raters 70.7% of the time, compared with 56.5% for an individual rater.
Benchmark accuracy and latency
Hume reports an average score of 0.82 across 10 leading benchmarks including MMAU, ParaPair and EmphAssess. Median latency is 0.23 seconds per screen, which Hume states is faster and more accurate than Gemini 3.8 Flash, Qwen3.8-Omni-Flash, Inkling-Small, gpt-audio-1.5, Gemini 3.5 Flash Lite and Nemotron 3 Nano Omni.
Realtime and batch use
The API supports realtime analysis and processing across hours of recordings. Questions can be added or changed at any time, with each five-second audio chunk answered independently.