On November 2 voice AI company Hume launched the Audio Decision API, which it calls the first audio-native model that answers questions about speech and sound. Developers upload raw audio, write each question as a sentence with candidate answers, and get back a probability for each answer, with no transcription step. It supports yes/no, multiple-choice, rating-scale and time-span questions, such as how many speakers there are, whether someone laughs, whether a caller sounds frustrated, or whether a siren is in the background. Per Hume, it was trained on more than 10 million human judgments and in an initial evaluation reached 70.7% agreement with a human panel's average judgment, versus 56.5% for an individual rater. Access is currently by request or waitlist.
Key Takeaways
- ✓Per Hume: 70.7% agreement with a human panel's average judgment in an initial evaluation, versus a 56.5% individual-rater baseline; sample size and languages were not disclosed.
- ✓Trained on more than 10 million human perceptual judgments (per Hume's X post); the blog cites tens of millions of listener ratings overall.
- ✓Four question types (yes/no, choice, scale, span); adding a question is one line of JSON, with no labeling or retraining.
- ✓The endpoint is
POST /v0/decisions/file, plus an input-token estimate endpoint, with Python and TypeScript SDKs. - ✓Hume says work that used to take 10+ models now takes one call; pricing has not been published.
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the problem
Voice agents and call-center QA usually transcribe a call first and then score the text. A transcript keeps only the words, so pauses, sighs, interruptions and tone, the signals that often explain why a call went wrong, are lost. Adding a new QA dimension typically means a new rubric, a new pipeline and new evaluations, taking days or weeks. Hume's Audio Decision API, launched November 2, lets developers question audio directly, the way they would prompt a language model.
Architecture and how it works
It is an audio-native model: raw audio goes in without transcription. Developers write each question as a sentence with candidate answers, and the API returns a probability for each answer that code can threshold, rank or combine. There are four question types: yes/no, choice, scale and span (for example, "where does someone laugh, sigh or gasp?"). One recording can be asked many questions at once, covering speakers, language and accent, emotional expression, background sound and conversation dynamics. The model builds on Hume's emotion-science research and large sets of human listener ratings.
Benchmarks
Per Hume, in an initial evaluation the model reached 70.7% agreement with a human panel's average judgment, versus a 56.5% baseline for an individual rater. Hume also says it beats existing audio methods on accuracy, speed and cost, but did not name the comparisons or give numbers. Sample size, languages and independent replication were not published.
Getting started
The endpoint is POST /v0/decisions/file, with an input-token estimate endpoint and run-history endpoints, plus the Python hume-audio-decisions and TypeScript @humeai/audio-decisions SDKs. Access is by request or waitlist, and a playground lets you upload a file and try questions. Good fits include scoring the naturalness of synthetic speech, giving a cascaded voice agent tone signals before it replies, and searching archives for calls where a customer sounded frustrated. Emotion judgments are perceptual, so calibrate thresholds against your own human-labeled samples before production. Pricing has not been published.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.