How Not to Develop an “umm” Detector Using Create ML

Speaking is difficult, especially if performed when others are within earshot. Many of us tend to use filler sounds like “umm” or “uhhh”, while our brains are processing the next thing we want to say. But, most of us wish we wouldn’t – especially the podcasters among us.

The phone in your pocket has a microphone and a neural engine. How hard could it possibly be to program it to give you a mild electric shock every time you say “umm”?

This talk will be a journey through a series of experiments of various attempts to create an Umm Detector. It will cover a number of Machine Learning-related topics ranging from Core ML to Apple’s easy-to-use speech recognition API, and the thought process behind using each method to detect “umm”s. The talk will also cover, step by step, the how to train a Machine Learning model using Create ML. By the end of the talk, you should have some idea of how NOT to create an Umm Detector and at least one way that works… at least… On My Machine™.

Also, machine learning is awesome… when used responsibly.