Hume Opens Its Audio Decision API to Answer Any Question About Sound
Hume's new Audio Decision API collapses a stack of specialized voice classifiers into one call that answers arbitrary questions about any recording.
- Hume launched the Audio Decision API, one model that answers arbitrary questions about audio with probabilities.
- Replaces 10+ specialized audio models (diarization, emotion, laughter, language ID) with a single call.
- Trained on 10M+ human judgments; agrees with human panels 70.7% vs 56.5% for a single rater.
- Averages 0.82 accuracy across 10 benchmarks at 0.23s median latency, beating Gemini 3.8 Flash and Qwen3.8-Omni-Flash per Hume.
- Works on realtime streams or multi-hour archives; questions can be yes/no, multiple choice, or scale.
- Waitlist only; lands as Hume sunsets EVI and TTS APIs on 13 November 2026.
Hume has opened the waitlist for its Audio Decision API, which uses one model to answer natural-language questions about audio. Developers define the possible answers and receive a probability for each one. Hume says the service supports live streams and recordings that run for hours.
According to the company, one API call can cover tasks that often require separate models for speaker diarization, emotion classification, language identification, laughter detection, and related audio analysis. The common probability output also gives applications a consistent interface for thresholds, routing rules, and human review.
One request, many judgments
The file endpoint accepts a multipart POST containing an audio file and a JSON object of named questions. Each question specifies an instruction, a response type such as yes/no, multiple choice, or scale, and optional criteria describing the available answers.
| Field | Purpose |
|---|---|
file |
The audio recording to analyze. |
questions |
A named collection of judgments to make. |
type |
The expected response format, such as boolean, choice, or scale. |
instructions |
The natural-language question presented to the model. |
criteria |
Optional names or descriptions for the answer choices. |
A file request can combine different question types and submit them with the same recording:
curl https://api.hume.ai/v0/decisions/file \
-H "X-Hume-Api-Key: $HUME_API_KEY" \
-F file=@support_call.wav \
-F 'request={
"questions": {
"frustration": {
"type": "bool",
"instructions": "Does the caller sound frustrated?"
},
"speakers": {
"type": "choice",
"instructions": "How many speakers are audible?",
"criteria": [
{"name": "One"},
{"name": "Two"},
{"name": "Three or more"}
]
}
}
};type=application/json'Because the questions are text, the endpoint can handle objective observations such as speaker count and language, perceptual judgments such as overlapping speech, and subjective ratings such as how reassured a patient sounds. Hume’s product demo returns a probability for each answer at five-second intervals; that cadence reflects the demo rather than a published service guarantee.
Claims behind the launch
Hume evaluates the model against aggregated human judgments. Its initial results report 70.7% agreement with a panel’s average judgment, compared with a 56.5% individual-rater baseline.
| Measure | Result | Meaning |
|---|---|---|
| Panel agreement | 70.7% | Model agreement with the panel’s average judgment. |
| Individual-rater baseline | 56.5% | Agreement achieved by an individual human rater. |
| Benchmark average | 0.82 | Average score across 10 company-selected benchmarks. |
| Median latency | 0.23 seconds | Median response time in Hume’s published comparison. |
| Training signal | More than 10 million judgments | Human judgments collected through Hume’s research. |
The agreement metric measures alignment with panel consensus. Tasks with objective labels, such as language identification or speaker count, require separate accuracy results.
Hume’s speed-and-accuracy chart places the Decision Model ahead of Gemini 3.8 Flash, Qwen3.8-Omni-Flash, Thinking Machines’ Inkling-Small, gpt-audio-1.5, Gemini 3.5 Flash Lite, and Nemotron 3 Nano Omni. The company has not disclosed the 10 benchmark datasets, so the aggregate score cannot yet be reproduced independently.
Signals transcripts leave out
Conventional voice-agent pipelines often send a speech-to-text transcript to a language model while omitting nonverbal cues from the waveform. Pauses, tone, laughter, overlapping speech, and background sounds can affect a conversation without appearing in the text.
Hume positions the Decision API as a parallel judgment layer whose probabilities can be combined with transcripts, application state, and business rules. Because the model receives the audio, it can score details represented by questions such as:
- Does the caller sound frustrated, sarcastic, or quiet?
- Are multiple people speaking at the same time?
- Did the agent sound empathetic during the past 10 seconds?
- Is a child or television audible in the background?
- Did the speaker switch languages?
From archives to live control
| Pattern | Use | Operational concern |
|---|---|---|
| Offline analysis | Filter recording archives for calls that warrant human review. | Threshold calibration, false positives, and review capacity. |
| Live routing | Feed probabilities into escalation logic or agent behavior. | Latency, window size, and smoothing across consecutive results. |
| Model evaluation | Score a fixed audio set against prompt-defined quality criteria. | Prompt versioning, model versioning, and reproducibility. |
Hume’s batch demo reports scanning 25,642 files at about 180 milliseconds per file, then filtering for recordings where the customer sounds frustrated and the issue remains unresolved. The reported throughput comes from the company’s demo and does not establish performance for recordings of different lengths or under production load.
Evaluation teams can add a quality dimension by writing another question, then rerun the same audio corpus as a voice model changes. Keeping the prompts, answer criteria, thresholds, and API version fixed is necessary for meaningful comparisons between runs.
Unknowns before production
- Availability and pricing: Access remains limited to a waitlist, and the product page does not publish pricing.
- Benchmark evidence: The reported results come from Hume, and the 10 benchmark datasets remain undisclosed.
- Operational terms: Production evaluation will depend on rate limits, supported codecs, data-retention and regional-processing terms, model versioning, streaming protocols, and service-level commitments.
- Product migration: Hume says its EVI and TTS APIs will close on , with associated account data permanently deleted afterward. The Decision API is a separate product line, and Hume has not described a direct migration path from those services.
The API offers a compact interface for audio judgments that otherwise require several task-specific models. Its production value will depend on independent accuracy tests, predictable latency, and commercial and data-handling terms that fit live and batch workloads.
Comments
