Hume Audio Decisions API answers a question most voice tools cannot: not what was said, but what was happening while it was said. Launched on November 2, 2026, it lets developers send one audio recording and ask plain-language questions about it — then get probabilities back, not a transcript.
That distinction is the product. A transcript tells you the words. Hume’s model listens to tone, interruptions, background sound, and speaker dynamics, and estimates how human raters would answer the same question.
What you can ask
Questions are written in plain language, and one request can carry several of them. The API supports four answer types:
- Yes/no. Did the agent interrupt the customer? Did the caller sound frustrated?
- Choice. Which speaker sounds more confident, or which emotion dominates the clip?
- Score. Rate frustration, warmth, or clarity on a scale you define.
- Time span. When exactly did the laughter, the sigh, or the overlap happen?
Inputs are one MP3 or WAV file per request, from 5 seconds to 5 minutes, heard as 16 kHz mono audio. The model returns no transcript. Answers are probabilities: Hume’s SDK defines a probability as the estimated share of human raters who would choose that answer.
Access is by waitlist, with a playground for real-time and file-upload experiments. Python and TypeScript SDKs cover every endpoint, and an input-tokens endpoint estimates the cost of a request before you run it.
The evidence: vendor numbers, labelled as vendor numbers
| Detail | What the source says |
|---|---|
| Launch | November 2, 2026; API access by waitlist, playground open |
| Input | One MP3 or WAV per request, 5 seconds to 5 minutes, 16 kHz mono |
| Price | $0.47 per million input tokens (audio, questions, context); answers not billed |
| Human agreement | 70.7% agreement with a human panel’s average judgment, vs 56.5% for an individual rater (Hume, initial evaluation) |
| Benchmark average | 0.82 across 10 benchmarks, per Hume’s product page (vendor claim) |
| Median latency | 0.23 seconds per screen, per Hume’s product page (vendor claim) |
| Training signal | More than 10 million human perceptual judgments, per Hume |
| Batch demo | 25,642 files scanned at about 180 milliseconds per file in Hume’s demo |
| Endpoints | POST /v0/decisions/file for files; run endpoints list past requests |
| Not a replacement | Hume’s text-to-speech and Empathic Voice Interface APIs end November 13, 2026; Decisions is a separate product |
Probability, not certainty. A 0.72 answer does not mean the event happened. It means Hume estimates that about 72% of human raters would answer that way. Build thresholds and human review around that meaning.
Turn call audio into a quality report
1. Pick five questions — Interruption, frustration, resolution, clarity, and background noise.
2. Score a sample batch — Run 100 calls and set thresholds from real distributions.
3. Flag, don’t auto-punish — Send low scores to a human reviewer first.
4. Track by agent and week — Trends beat single-call verdicts.
5. Keep the transcript too — Hume gives judgments, not words; pair both.
How to profit from it
Call quality is a service small businesses already buy, usually by listening to a sample by hand. Hume automates the listening part:
- Sell a call-quality audit. Score a client’s last 500 support calls on interruption, frustration, and resolution, and deliver the worst 20 with timestamps.
- Offer voice-agent evaluation. Teams building voice agents need to know whether the agent talks over people. That is a question Hume answers directly.
- Build audio search. Find every call where a customer sounded angry and the issue stayed unresolved — a filter no transcript search can do alone.
- Price per insight, not per minute. At $0.47 per million input tokens with answers unbilled, your margin is the report, the thresholds, and the coaching that follows.
- Do not oversell accuracy. The 70.7% figure is Hume’s initial evaluation, with sample size and languages undisclosed in the coverage we read. Present it as a screening tool with human review.
Sources
Hume’s launch as covered by Decision Models, S3DAT News, and Superpower Daily. Benchmark and agreement figures are Hume’s own claims; join the waitlist and test the playground before promising a client a number.
Frequently Asked Questions
What is the Hume Audio Decisions API?
An API launched November 2, 2026, that answers plain-language questions about raw audio with probabilities, without producing a transcript.
What questions can it answer?
Yes/no, choice, score, and time-span questions — for example, whether a caller sounded frustrated, or when an interruption happened.
What audio can I send?
One MP3 or WAV file per request, from 5 seconds to 5 minutes, heard as 16 kHz mono audio.
How much does it cost?
$0.47 per million input tokens (audio, questions, and context); answers are not billed, per Hume’s billing docs.
Is it generally available?
No. API access is by waitlist, with a playground for real-time and file-upload experiments.
How accurate is it?
Hume reports 70.7% agreement with a human panel’s average judgment in an initial evaluation, vs 56.5% for an individual rater. That is a vendor figure, not independent verification.
Does it replace Hume’s voice generation APIs?
No. Hume’s text-to-speech and Empathic Voice Interface APIs end November 13, 2026, and Decisions is a separate product with no described migration path.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.