Hume Audio Decision API: waitlist access, audio limits and billing
Hume introduced its Audio Decision API on November 2, 2026. It returns judgments about recordings. The product page currently directs applicants to a waitlist, so the announcement is not evidence of unrestricted API availability.
What the documented endpoint accepts
The overview and limits guide describe a file-based request. Marketing demonstrations of live conversations should not be treated as proof of a generally available streaming endpoint.
| Integration point | Documented requirement |
|---|---|
| Audio | One MP3 or WAV, 5 seconds to 5 minutes |
| Total request size | 25,000,000 bytes, including JSON and audio |
| Question types | Yes/no, choice, score and time span |
| Question allowance | Up to 20 yes/no, choice and score questions combined, plus 5 span questions |
| Request rate | 60 per minute per organization, shared across endpoints and keys |
A five-minute WAV can exceed the byte limit. The service processes audio as 16 kHz mono. Longer recordings need separate qualifying segments.
Answer probabilities estimate how human raters would respond. A value near 0.5 indicates disagreement among listeners, not necessarily model uncertainty. Keep question sets and order consistent when comparing recordings. Changing them can affect answers. Model updates can also change results without an API-version change.
Billing is documented, but public rates are not established
The billing guide charges successful requests for input tokens from audio, questions and context. Answers are not billed. Silence counts toward audio duration, and overlapping segments send the same audio repeatedly, increasing usage.
Token estimates and run queries are unbilled, as are requests rejected with 4xx errors. Estimates still count toward the request limit. The documentation sends readers to their account’s Plan page or sales for prices. ByteForward has not verified a public numerical rate. A waitlist announcement should not be read as a free-service offer.
Treat listener judgments as measurements to validate
Hume reports 70.7% agreement with panel judgments versus 56.5% for individual raters. This vendor comparison does not establish objective knowledge of internal emotions.
A useful pilot would separate observable audio events, such as overlapping speakers, from subjective impressions. Review false alarms, differences across recording conditions and the consequences of acting on a score. Keep uncertain cases available for human review rather than letting a single threshold decide consequential outcomes.
ByteForward has not submitted recordings or tested the API.
Watch Hume’s official product introduction. This vendor demonstration is not an independent evaluation.
Original editorial illustration by ByteForward, generated with AI. Not a product interface or measured audio result.
