Hume launches the first decision model for voice AI

Hume launched an Audio Decision Model that returns a probability for each possible answer to a developer-submitted question about a recording, and the company reports it agrees with a panel of human raters 70.7% of the time versus 56.5% for a single human rater. Hume said the model averages 0.82 across 10 benchmarks including MMAU, ParaPair and EmphAssess at 0.23 seconds median latency per screen, scoring higher and answering faster 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. Hume cited 1B client hours processed and more than 10M human judgments across its work, and positions the Decision Model API as a decision layer in voice-agent systems alongside speech recognition and generation.

Decision models for voice AI The bar is human. Hume diagnoses system breaks and feeds failures into improvement. A sample telehealth call plays while three questions about it are answered for every five seconds of audio. Why audio needs a decision model Voice AI considers tone, pauses, and sounds. Developers set Q&A; the audio decision model gives probabilities for quick audio judgments. Different questions. The same audio. Ask objective, subjective or perceptual questions about speech and sound. Select a tab to explore examples. One speaker Two speakers Three or more speakers Find the audio worth acting on Ask more of the same audio Apply several questions to the same recordings, from naturalness and clarity to empathy and timing. Designed for fast evaluation Repeat evaluations across your collection as your model or criteria change. Pinpoint cases for your next step Find candidate examples for training, isolate evaluation failures, and surface cases for escalation review. More reliable than a single listener The Audio Decision Model agrees with the human crowd 70.7% of the time, while a single human rater agrees 56.5% of the time. - Number of client hours processed. - 1B hours - Human judgments across Hume’s work. - 10M+ Agreement with a panel of human raters: Audio Decision Model 70.7%, Individual Rater 56.5%. Accuracy against speed: Hume’s Decision Model averages 0.82 over 10 leading benchmarks at 0.23 seconds median latency per screen, scoring higher and answering faster 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. The Audio Decision Model answers audio questions faster and more accurately than any other model on the market. - Average score over 10 leading benchmarks: MMAU, ParaPair, EmphAssess, etc. - 0.82 - Median latency per screen - 0.23 s Compare the judgments that matter Accuracy across the criteria you care about. 1. Hume’s Decision Model Hume82% 2. Gemini 3.8 Flash Google78% 3. Qwen3.8-Omni-Flash Alibaba76% 4. Inkling-Small Thinking Machines67% Questions about the model It’s Hume’s audio-native decision model. Developers submit audio, a question and possible answers; it returns a probability for each answer. It helps teams evaluate speech and sound, assess how a voice comes across to listeners, and inform decisions in voice applications. Audio decisions that fit your stack. Send audio, your question and possible answers. The audio decision model returns a probability for each Use the Decision Model API as a decision layer in a voice-agent system with speech recognition and generation.