How would you test whether a machine really understands a cat?

by•

My name is Ano Wen, and I co-founded a pet AI company in Seattle. Last year our team published a paper at CogMAEC '25, a workshop held with ACM Multimedia 2025, and it started with a question I still think about: when a cat meows, what would it take for a machine to do more than label the sound?

Most earlier work in this space stopped at classification: a sound goes in, a category comes out. In our study, we classified cat vocalizations into ten categories and reached 92.09% accuracy on a dataset of 2,961 cat sounds, evaluated with 10-fold cross-validation. But a label alone felt like a dead end, because a label is not a conversation. So the system also factored in the individual cat's personality and recent interaction history, and generated a response, spoken from a small wearable prototype and shown as text in an app, in under three seconds end to end.

The part I keep coming back to is the evaluation problem. There is no ground truth for what a cat means. Accuracy on a labeled dataset only tells you the classifier agrees with human annotators. It does not tell you the cat was understood.

So I am curious how others here would approach this. If you work on animal communication, affective computing, or voice interfaces for users who cannot speak for themselves, how do you evaluate a system like this? What evidence would convince you that it works?

The paper, for context:

9 views

Add a comment

Replies

Be the first to reply

Have a question or a thought to share? Add a comment above to start the conversation.