The short version
- Hume AI's Audio Decision API is pitched as one call that answers human-style questions about audio.
- The company says it was trained on more than 10 million human judgments and works in realtime or across hours of recordings.
- There is still no public build, download, or official page; Hume is taking access requests after a short preview.
At a glance
| Maker | Hume AI |
|---|---|
| Announced | November 2, 2026 |
| Training | 10M+ human judgments |
| Access | Request only; no public build yet |
| Category | AI & Models |
| Creator / publisher | Hume AI |
| Heat | 🔥 2.3M Heat |
| Announced | |
| Data updated |
One call instead of a model stack
What used to take 10+ models is now one call.
Hume AI introduced the Audio Decision API as a way to ask a recording a direct question and get an answer a human listener might give. The launch examples are concrete: how many speakers are present, whether anyone is laughing, and whether the caller is frustrated.
The pitch is consolidation. Hume says work that previously required more than 10 separate models now fits in a single call, and that the system can run in realtime or across hours of recordings.
- Speaker count
Ask how many people are speaking in the audio.
- Laughter
Ask whether anyone is laughing.
- Caller mood
Ask whether a caller sounds frustrated.
- Open questions
Hume frames the interface as ask anything, with those three as the published examples.
Judgments trained to match human listeners
Trained on 10M+ human judgments, so it answers the way human listeners would.
The only training detail Hume has published is scale and target: more than 10 million human judgments, with answers meant to line up with how people actually hear a clip. That is a decision layer on top of audio, aimed at counts, events, and caller state rather than a transcript alone.
No parameter count, dataset breakdown, latency figure, or benchmark table has been released with the announcement. The operating range that is stated is time: realtime on one side, and hours of recordings on the other.
| What Hume has stated | What remains unpublished |
|---|---|
| One call replaces a stack of 10+ models | Which earlier models it stands in for |
| Trained on 10M+ human judgments | Model size, architecture, and eval scores |
| Realtime or hours of recordings | Latency, file limits, and pricing |
| Request access | A public app, API docs page, or open weights |
Where a listening decision actually helps
The published examples point at contact centers and any workflow that already records conversations. A team could ask whether a caller is frustrated, whether laughter shows up in a session, or how many speakers are in a clip, without chaining separate detectors for each question.
Long recordings are part of the claim, so review of hours of audio sits next to live calls. Until access opens, those stay intended uses from the announcement, demonstrated so far in a short preview.
Access, pricing, and what is still open
Request access below.
Hume has invited people to request access. No public build, download, or official page has been published yet, and the creator has only shown the product in a short preview. There is no self-serve app, documented API, or open-weight release to install.
Price, rate limits, supported formats, and how answers are returned are not in the announcement. The comparable-product story is Hume's own: one call in place of more than 10 models, scored against human listening judgments, with independent benchmarks still ahead.
FAQ
Can you use the Hume Audio Decision API today?
You can request access. Hume has not published a public build, download, or official page, so there is no open app or weight file to run on your own.
Did Hume publish a price or model size?
The announcement covers the one-call design, training on more than 10 million human judgments, and realtime or multi-hour audio. It does not list a price, rate limits, or model size.
