Hume Audio Decisions: test one detail in your voice feature

Choose a detail you can hear

If you are building a voice feature, start with one detail that makes it easier to use. Do two voices speak at the same time? Does background music cover the words? Prepare a few approved recordings and write down what you hear before asking AI to check them.

On November 2, Hume introduced its Audio Decision API, a connection that lets an app send a recording and questions to its service. It can help a builder compare voice features. Access is currently by request through a waitlist.

Read the announcement: Hume: Introducing the Audio Decision API.

Understand what the answer means

You provide the question and the possible answers. Hume returns numbers estimating how a panel of human listeners would respond. It does not return a transcript, a spoken reply or an explanation.

For a yes-or-no question, a result of 0.8 means the model estimates that 80% of the panel would answer yes. It does not prove the answer is true. A result near 0.5 suggests listeners would split.

Keep listening to the samples yourself. Use the answers to support a review, rather than treating a number as a complete judgment of a person or conversation.

Check access and the cost first

The current connection accepts one audio file lasting five seconds to five minutes per request, within its file-size limit. Longer recordings need to be split. The launch does not mean every account can immediately connect an app.

Hume currently lists US$0.47 per million input tokens, the units it counts from the audio and text you send. Recording length makes up most of the charge, including silence. Questions and any description you add also count; the answers are not billed.

An organisation receives a one-time trial allowance of one million input tokens after access is granted. The account shows when the credit expires. Ask your builder to use Hume's free estimate tool to price your sample job before sending the audio. These are service charges, not the complete cost of a finished app.

Practise one check you can repeat

You can prepare the exercise while waiting for access. Use recordings you are allowed to use, such as a staged conversation with everyone's agreement. Keep private customer information out of a practice set.

1. Choose one question, such as whether two voices speak at the same time, and define what counts as yes.
2. Listen to several short clips, including a clear example and an awkward one. Record your expected answers.
3. Once access is approved, ask the same questions in the same order for every clip and compare the results with your listening.
4. Note missed cases, disagreements and the cost. Decide which results need a person's review before using the check in your testing process.

Make the check part of your build

Hume's standard terms describe internal development and support for your own apps, and restrict offering the service to others. Confirm what your account agreement permits before planning a connection that customers would use.

Before sending real recordings, review Hume's terms and privacy policy with whoever owns your information handling. Confirm what you have permission to upload and how the service will handle it.

Our perspective is that a useful voice feature starts with knowing what good sounds like. Bring your examples into a training session or build discussion, and repeat the check when your voice feature changes.