What happened
Hume’s product page introduces the Audio Decision Model API as one model for asking questions directly about audio. A developer submits a recording, a question, and, where applicable, candidate answers; the model returns a probability for each answer. The page distinguishes factual questions—such as how many speakers are audible, what language is spoken, or whether there is background speech—from listener judgments and interaction decisions. Examples include whether a patient sounds anxious, how a clinician’s tone is best described, and how reassured a patient sounds at the end of a consultation.
The API is presented as configurable rather than tied to a fixed label set. The page says developers can add or change questions, and that several questions can be applied to the same recording. Its example request sends a WAV file and question definitions to https://api.hume.ai/v0/decisions/file, with question types including a yes/no-style “noul” answer and a choice among named criteria. The product page also illustrates processing a telehealth call in five-second chunks, with three questions answered for each chunk. This is an example of the interface, not a stated universal chunking requirement.
Hume describes several intended uses: repeating evaluations across a recording collection as a model or criteria change, finding candidate examples for training, isolating evaluation failures, and surfacing cases for escalation review. A demonstration shows searching a collection of 25,642 files for frustrated customer calls and reports a scan rate of 180 milliseconds per file. The page says the model is designed for fast evaluation, but does not explain the operational setup behind that demonstration.
The company reports 70.7% agreement between its model and a panel of human raters, compared with 56.5% agreement for an individual human rater. It also reports an average score of 0.82 across ten benchmarks and median latency of 0.23 seconds per screen. Hume’s comparison lists Gemini 3.8 Flash at 78%, Qwen3.8-Omni-Flash at 76%, and Inkling-Small at 67%; the page also names gpt-audio-1.5, Gemini 3.5 Flash Lite, and Nemotron 3 Nano Omni in the benchmark comparison. These are Hume’s reported figures. The page offers API documentation and a waitlist/request-access path, but gives no public price, quota, or general availability date.
What the archive adds
This is a new product event in a developing decision-model category. Our earlier OpenAI Decisions API coverage described a general-availability API with gpt-6-luna support and up to 400 questions per request. Hume’s announcement differs in focusing specifically on audio input and audio judgments; it is not evidence that the two APIs share an implementation or access model.
Trend signal
This supports the pattern that model outputs are becoming task interfaces: a builder supplies bounded questions and answer choices rather than receiving only open-ended prose. The relevant pattern is Model answers become task interfaces. Hume’s API makes those interfaces audio-native, although the product page does not establish how reliably probabilities transfer across domains or recording conditions.
Why it matters
For teams building voice agents, the API suggests a way to turn recordings into structured evaluation signals without first reducing every judgment to a transcript. Questions about interruption, background speech, tone, or apparent frustration may capture properties that words alone miss. Re-running the same criteria over a corpus could help teams find calls for review, compare system versions, or assemble examples for improvement.
The practical value depends on validation. A probability is not itself a verified ground truth, especially for subjective states such as anxiety or reassurance. Teams would need to define what decisions the score can safely influence, compare results against their own raters and populations, and route consequential or ambiguous cases to humans. The product page describes an evaluation and triage layer, not a replacement for clinical or customer-service judgment.
Limits and uncertainty
The source is Hume’s own product page, and the accuracy, latency, human-agreement, and scan-rate figures are vendor claims. It does not provide benchmark protocols, datasets, confidence intervals, independent replications, or enough detail to assess whether the comparisons use equivalent conditions. The page also does not specify supported audio formats beyond the example WAV upload, maximum recording length, throughput limits, data retention, privacy controls, regional availability, pricing, or service-level commitments. API access is presented through a request/waitlist flow, so the page does not establish broad availability. It remains unclear how probabilities should be calibrated for a particular application and how well subjective judgments hold across accents, languages, or recording quality.
Reader takeaway
Hume is packaging questions about sound and speech as a configurable API that returns probabilities, with evaluation and triage as its clearest use cases. Builders can treat it as a candidate measurement layer for voice systems, but should not treat its headline metrics as independent proof: access, costs, limits, and application-specific reliability still need to be established.