Hume AI has launched its Audio Decision API, a new tool that allows developers to ask questions directly about speech and other sounds in audio recordings. Instead of relying only on transcripts, the API analyses raw audio and returns structured answers with probabilities, helping applications identify emotional expression, speaker behaviour, background sounds and conversational events.
Announced on November 2, 2026, the API is designed for applications such as customer-service agents, voice AI evaluation, call-centre analytics and audio search. Developers can ask whether a customer sounds frustrated, whether an AI-generated voice sounds natural, or when someone laughs or interrupts another speaker. Hume says its model achieved 70.7% agreement with the average judgment of a human evaluation panel, compared with 56.5% for an individual human rater in its initial evaluation.
Key Takeaways
- Hume AI launched its Audio Decision API on November 2, 2026.
- The API analyses raw audio instead of depending exclusively on transcripts.
- Developers can ask custom questions about emotion, speech, speakers and background sounds.
- Responses include probabilities and structured results that applications can use.
- The API supports yes/no, multiple-choice, rating-scale and timestamp-based questions.
- Hume reports 70.7% agreement with the average judgment of a human panel.
- The company says its model scored 0.82 across 10 benchmarks, with approximately 0.23-second median latency in its reported comparison.
- The API can support customer-service analytics, voice-agent evaluation and large-scale audio search.
- Developers can request access through Hume’s waitlist and experiment through its playground.
What Is Hume AI’s Audio Decision API?
Most speech AI systems begin by converting audio into text. A separate language model then analyses the transcript to understand what happened in the conversation.
However, transcripts do not capture everything.
Consider a customer telling a support agent, “That’s fine.”
The words appear positive or neutral, but the speaker may sound frustrated, disappointed or uncertain. A transcript alone cannot reliably preserve these vocal characteristics.
The same problem applies to pauses, interruptions, laughter, sighs, background noise and changes in speaking style.
Hume’s Audio Decision API is designed to analyse these signals directly from audio.
Developers provide an audio recording and write questions in ordinary language. The model returns probabilities for the possible answers rather than generating a long textual explanation.
This makes the output easier to connect to software applications, automated workflows and AI agents.
How Does the API Work?
The API allows developers to define the questions they want answered and the possible response formats.
It supports four main question types:
| Question type | What it does | Example |
|---|---|---|
| Yes/No | Estimates whether something is present | Does the customer sound frustrated? |
| Choice | Selects between defined options | Which emotion best describes the speaker? |
| Score | Rates audio against an ordered scale | How natural does the voice sound? |
| Span | Identifies when an event occurs | At what time does someone laugh? |
Developers can ask several questions about the same recording in one request.
For example, a customer-service company could analyse a call to determine whether the customer sounds frustrated, whether the agent interrupts them, how many people are speaking and whether background noise affects the conversation.
The API returns probabilities that developers can use to set thresholds and trigger business rules.
A company could, for instance, flag calls for human review when the probability of customer frustration crosses a predefined threshold.
Importantly, these are estimates of how human listeners might judge the audio. They should not be treated as definitive measurements of someone’s internal emotional state.
Why Analysing Audio Beyond Transcripts Matters
Traditional speech recognition focuses primarily on identifying spoken words.
But human communication involves much more than language.
The way someone speaks can change the meaning of the same sentence. A pause might indicate uncertainty, while an interruption could suggest that a conversation is becoming difficult.
Hume’s model is designed to capture several categories of information:
- Emotion and expression: Frustration, empathy, excitement and other vocal characteristics.
- Speech and language: Language identification and speech-related characteristics.
- Speakers and voices: Speaker-related information and voice characteristics.
- Conversation dynamics: Interruptions, turn-taking and overlapping speech.
- Acoustic environment: Background noise and other environmental sounds.
- Non-speech events: Laughter, sighs, gasps and other audible events.
- Intent and outcomes: Questions about what a speaker is trying to accomplish.
The company says developers can also update questions as their requirements change without retraining the model or building a separate classification pipeline.
This could make it easier for businesses to adapt their audio-analysis systems when they introduce new quality standards or customer-service requirements.
Hume Reports 70.7% Agreement With Human Evaluators
One of the most important parts of the announcement is Hume’s evaluation of its model against human judgments.
The company says the Audio Decision Model achieved 70.7% agreement with the average judgment of a human evaluation panel.
For comparison, an individual human rater achieved 56.5% agreement with that panel average.
<text weight=”medium” color=”default”>Reported agreement with average human-panel judgment</text>
| Evaluator | Agreement |
|---|---|
| Hume Audio Decision Model | 70.7% |
| Individual human rater baseline | 56.5% |
<Chart content={{“chartType”:”bar”,”meta”:{“description”:”Figures reported by Hume in its initial evaluation. Agreement is not the same as universal accuracy.”},”xKey”:”evaluator”,”series”:[{“dataKey”:”agreement”,”label”:”Agreement”,”valueSuffix”:”%”}],”yAxisMin”:0,”yAxisMax”:100,”data”:[{“evaluator”:”Hume model”,”agreement”:70.7},{“evaluator”:”Individual rater”,”agreement”:56.5}]}}/>
Hume says the model was trained using years of emotion-science research and millions of human judgments about how people perceive speech and sound.
The company also compared its model with other audio-processing systems across 10 benchmarks.
Its reported average score was 0.82, with approximately 0.23 seconds of median latency in the comparison. Hume says this performance was ahead of the competing systems included in its evaluation.
These are company-reported results, not an independent guarantee that the model will outperform every competing system in every real-world environment. Performance can also depend on the recording, question and evaluation criteria.
How Businesses Could Use the Technology
The API could be particularly useful in industries where conversations need to be evaluated at scale.
1. Customer-Service and Contact Centres
Contact centres record thousands of customer interactions.
Traditional quality-assurance systems often rely on transcripts, predefined keywords and manual reviews.
Hume’s API could help identify conversations involving frustration, interruptions, excessive background noise or poor communication quality.
Businesses could use these signals to prioritise calls for human review and improve customer-service workflows.
2. Testing AI Voice Agents
Companies are increasingly deploying voice-based AI assistants for customer support, bookings and other services.
A voice agent may provide the correct information but still sound unnatural or fail to respond appropriately to a customer’s emotional expression.
Developers could use the API to evaluate generated speech, measure how natural a voice sounds and compare different model versions.
These evaluations could also help teams improve their voice models during development.
3. Searching Large Audio Libraries
Another potential application is searching large collections of existing recordings.
Hume says its technology can make an archive containing 40,000 audio recordings searchable in minutes.
For example, a company could look for every recorded call where a customer sounded frustrated or where a particular sound occurred.
This could reduce the need for extensive manual listening, transcription and labelling before analysing audio collections.
4. Building More Responsive Voice Agents
A voice agent could combine the transcript with Hume’s audio judgments before deciding how to respond.
If a customer says they are satisfied but sounds uncertain or frustrated, the application could use that additional signal to decide whether to ask a clarifying question or escalate the interaction.
The API provides an additional layer of information; the application or language model still needs to interpret the results and decide what action to take.
Is the API Available to Developers?
Hume has announced the Audio Decision API as available and provides documentation, Python and TypeScript SDKs.
Developers can request API access through the company’s waitlist and experiment with their own questions using its playground.
The documented file endpoint accepts MP3 and WAV recordings between five seconds and five minutes long. It returns structured answers, probabilities and information about the request.
There is an important distinction between Hume’s real-time ambitions and the currently documented API workflow. The company describes rapid analysis and a real-time playground experience, but its documented file endpoint is not the same as a fully documented continuous streaming API. Hume also describes broader live and ambient applications as a longer-term opportunity.
The Bigger Picture
The launch reflects a growing shift in voice AI development: understanding how something is said is becoming as important as recognising the words themselves.
As AI assistants become more common in customer service, healthcare administration, sales and other industries, developers need systems that can interpret vocal cues and conversational context.
Hume is positioning its technology as a decision layer that can sit alongside speech recognition and large language models. Rather than replacing those systems, it provides structured information about the audio that they may otherwise miss.
The wider opportunity is to make voice applications more responsive and improve how businesses evaluate human and AI conversations.
Looking Ahead
The next challenge for Hume will be demonstrating how reliably its model performs across different languages, accents, recording conditions and real-world conversations. Independent evaluations will also be important, particularly for applications involving emotional judgments that can vary between listeners and cultures.
For developers, the Audio Decision API offers a new way to build applications that understand more than transcripts. If its reported performance holds up in practical deployments, it could become a useful component in customer-service analytics, voice-agent testing and large-scale audio intelligence systems.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
