The Pulse
Hume Ships Audio Decision API for Direct Speech Analysis
Hume’s Audio Decision API lets developers ask questions about speech and sound and receive probabilities for answers they define. The company reports evaluation and benchmark results, while access is available by waitlist.
AI.info Team · Written with AI · fact-checked ·
“This is a different architecture. Raw audio goes in, an answer with a confidence score comes out, structured and ready for your code or your agent to act on.”
Hume AI Team
Hume introduced its Audio Decision API on November 2, pitching it as a way for software to judge speech and sound directly rather than depend on transcripts alone. Developers define questions and possible answers; the API returns probabilities that an application can use in its own rules or workflows. Hume says the model can assess cues such as frustration, overlapping speakers, laughter and background sirens.
Questions go straight to the recording
The API accepts questions in plain language and supports four answer formats: yes or no, a choice among options, a rating scale, or a time span locating an event in a clip. A developer could ask whether a caller sounds frustrated, which tone best describes an agent, or where someone laughs or sighs. Hume says several questions can be applied to the same recording, with confidence thresholds set to connect results to business rules.
That design targets a gap in transcript-based call analysis. A transcript records words, but not a pause, sigh or the moment two people speak over one another; Hume argues those cues can help explain how a conversation is going. In its customer-service example, the company suggests combining audio judgments with transcription and an LLM, which could use both inputs to decide whether a voice agent should wait, clarify or escalate.
Hume reports 70.7% agreement with panel judgments
Hume says tens of millions of listener judgments on emotional expression, vocal delivery, naturalness and conversational fit informed the model. In an initial evaluation, the Audio Decision Model agreed with the average judgment of a human panel 70.7% of the time, compared with 56.5% for an individual-rater baseline. The launch post does not describe the evaluation set or provide enough methodological detail to assess how broadly those results apply.
The company also reports an average score of 0.82 across 10 audio benchmarks and median latency of about 0.23 seconds. Its comparison image lists Gemini 3.8 Flash, Qwen3.8-Omni-Flash, Inkling-Small, gpt-audio-1.5, Gemini 3.5 Flash Lite and Nemotron 3 Nano Omni. Those figures are company-reported results, not an independent assessment.
Waitlist access limits the launch
Hume’s announcement directs developers to request API access through a waitlist, while a playground offers real-time and file-upload tests. The company describes use cases ranging from evaluating generated voices to searching call archives for particular sounds or conversational patterns. It says an archive of 40,000 recordings could become searchable in minutes, though the post does not provide the conditions behind that estimate.
The company says questions can cover speech and languages, speakers and voices, emotional expression, acoustic environments, non-speech events and conversation dynamics, among other areas. Developers can ask multiple questions about the same recording, update questions in real time and apply them in bulk across audio libraries. Hume advises teams evaluating voice models to validate their assessments against representative human ratings.