HYPE OR FACT? · CHECKED 2 NOVEMBER 2026
Hume's Audio Decision API: "10+ models, now one call", trained on 10M+ human judgments
OUR VERDICT
Partly factOur original verdict: “Fact or hype: partly fact. The API is real and well documented, but the "10+ models" line is marketing, the human-judgment and benchmark numbers are Hume's own, and access is by waitlist.”
THE CLAIM
“What used to take 10+ models is now one call. Trained on 10M+ human judgments, so it answers the way human listeners would. Realtime, or across hours of recordings.”
What’s actually true
- The product is real. Hume's docs went live on 2 Nov with an endpoint, Python and TypeScript SDKs, limits and billing rules. You send one audio file plus up to 25 plain-language questions and get a probability for each allowed answer.
- Access isn't open. The launch post and product page say to join a waitlist to request API access, and there's no public price. Billing is by input tokens, mostly the length of the audio, with plans shown only after you sign in.
- '10+ models into one call' is a framing, not a measured fact. Hume's own diagram shows five separate steps it replaces (speech-to-text, speaker ID, expression, language ID and AI-voice detection). Asking many questions in one call is real, but we found no list of ten models.
- '10M+ human judgments' is Hume's own figure. Its product page says 10M+ judgments across Hume's work, and the blog says tens of millions informed its VoiceEQ benchmark and helped train the model. We can't verify the number from outside.
- Hume says the model agreed with a human panel 70.7% of the time, against 56.5% for a single rater, and scored 82% across 10 benchmarks, ahead of Gemini 3.8 Flash (78%) and Qwen3.8-Omni-Flash (76%). These are Hume's own initial tests. Outside coverage (The Deep View, FourWeekMBA) repeats them, and FourWeekMBA notes they haven't been independently tested.
- 'Realtime, or across hours of recordings' is stretched. Each request takes one file of 5 seconds to 5 minutes, so hours of audio means cutting it into pieces. The playground answers live audio every 5 seconds, but the blog says live and ambient use 'remain a longer-term opportunity'.
- We haven't watched the 205-second launch video. This check uses the post, Hume's website, blog and docs, and press coverage.
Sources
Cost
No public price: billed by input tokens (mostly audio length), plans shown after sign-in. Waitlist access.
USE IT IF
Contact centres, voice-agent builders and teams testing text-to-speech who need to judge how calls and voices sound, at scale.
SKIP IT IF
Small businesses wanting a ready-made app: it's a developer API behind a waitlist with no public price. Also not for judging staff or customers' emotions without checking consent and privacy rules.
Related AI tools
Related
How we knowChecked on 2 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗