Teachable Machine is a free, browser-based tool from Google that lets you train a working machine-learning model — for images, sounds, or body poses — in minutes, using nothing but your webcam or microphone and no code at all. For artists and makers building interactive work, it’s the single fastest way to get a custom recognizer: a project that knows the difference between an open hand and a fist, a person present or absent, or one object versus another. This is a getting-started guide to training your first gesture recognizer and putting it to use.
Watch: How to Use Google’s Teachable Machine: Step-by-Step for Beginners (YouTube)
Step 1: Gather examples into classes
Go to the Teachable Machine site and start a new Image Project (the Standard model is fine). You’ll create classes — the categories you want the model to tell apart — and give each one examples. For a simple gesture recognizer, make one class for “open hand,” one for “fist,” and one for “nothing” (an empty background). Hold the pose in front of your webcam and record a few hundred sample frames per class, varying the angle, distance, and lighting a little so the model generalizes instead of memorizing one exact position. That “nothing” class matters more than beginners expect — without it, the model will force everything into a gesture.
Step 2: Train and test, then fix the gaps
Click Train Model and wait — training happens right in the browser, and it’s quick. When it finishes, the preview runs live: hold up each gesture and watch the confidence bars. This is where you actually learn how ML behaves. If it confuses two classes, the fix is almost always more and more varied data — different lighting, backgrounds, hand sizes, positions. Retrain and test again. A few rounds of “test, find the failure, add examples for it, retrain” will teach you more about machine learning than any lecture, and it’s the core skill the tool exists to build.
Step 3: Export and connect it to something
A recognizer is only interesting once it drives something. Teachable Machine can export your trained model — most usefully as a TensorFlow.js model you can drop into a web page, or via a hosted link. From there it pairs naturally with p5.js (through the ml5.js library) so your sketch can react to gestures: change color on an open hand, trigger a sound on a fist, advance a visual when someone appears. That handoff — trained model into a creative-coding sketch — is the bridge from “cool demo” to “actual interactive installation.”
Where to take it
Once image gestures click, try the Pose project (which tracks body keypoints — great for whole-body, movement-based installations) and the Audio project (train it on claps, words, or sounds). The honest limits are worth knowing: Teachable Machine is a prototyping tool, its models are relatively simple, and it won’t match a purpose-built pipeline for accuracy or robustness. But for turning an idea for a gesture-controlled piece into a working prototype the same afternoon, nothing is faster.