Speech in Noise part 2: Why is Computational Cost Such a Problem for AI?

Posted by Dr Andrew Simpson on Nov 14, 2020 6:50:23 PM

What’s the easiest way to make an AI perform better? The first answer is usually: make it bigger.

But if your goal is AI for real-time speech enhancement that needs to run on a small hearing device platform, making your AI bigger is the exact opposite of what you want to do. And if the best way to improve your AI is to make it bigger, then the best way to make your AI worse is to make it smaller.

This is another example of the uncertainty principle for real-time speech-enhancement AI. You can have it big and sounding good, or you can have it small and sounding bad, but you can’t have both.

At Chatable, we are developing the world’s first real-time zero-latency AI for speech in noise that can run on-device.

Keep an eye out for updates.