Teachable Machine

What is Teachable Machine?

Teachable Machine is a free web tool from Google that trains small machine learning models to recognise your own images, sounds and body poses, with no coding required.

The workflow has three stages. You gather examples from a webcam, microphone or uploaded files and sort them into labelled classes. You then press a train button, and the tool fits a model inside your browser using TensorFlow.js, Google’s JavaScript machine learning library. Finally, a live preview lets you test the result before you export it for a website, an app or a hardware project.

Google Creative Lab built the tool, and version 2.0 was announced in November 2019, according to VentureBeat coverage at the time. Its open-source community repository on GitHub carries an Apache 2.0 licence and describes the project as an experiment rather than an official Google product. The training page still loads on desktop Chrome or Safari, but the helper libraries on npm have not had a new release since August 2021, so treat it as stable but lightly maintained.

Key Features

Flexible exports matter most here, because a model trained in minutes can move straight into a web page, an Android app, a Coral board or an Arduino.

• Three project types: You can train image, audio or pose models from the same start page. Each type uses its own helper library, so a webcam gesture project and a sound classifier follow the same simple steps.
• In-browser training: Training runs on your own computer through TensorFlow.js rather than on a remote server. Coverage of the 2.0 launch reported that samples stay on the device unless you choose to save them to Google Drive.
• Live testing preview: A preview panel shows class confidence in real time as soon as training finishes. You can point a webcam or speak into a microphone and see straight away whether the model confuses two classes.
• TensorFlow.js export: Every project type, including pose, can export to TensorFlow.js with ready-made JavaScript snippets. Image and audio projects also come with p5.js examples, and the export panel can supply a hosted model link for quick prototypes.
• TensorFlow Lite and Keras: Image models export to Keras as an .h5 file with Python sample code, plus an OpenCV variant. Image and audio models also export to TensorFlow Lite for Android, and image models to Coral Edge TPU.
• Embedded image model: A smaller embedded image option targets microcontrollers through TensorFlow Lite for Microcontrollers. The official guide pairs it with an Arduino Nano 33 BLE Sense and an OV7670 camera module.

Pros

Zero cost and zero setup make Teachable Machine the easiest first step into machine learning for a classroom or a hobby project.

• Completely free: No paid plan, credit system or subscription appears anywhere in the tool. Schools can use it with whole classes without budgeting for licences or tracking usage limits.
• No account to start: You can open the training page and begin collecting examples straight away. Signing in is only needed if you want to save a project to Google Drive for later.
• Privacy-friendly design: Because training happens in the browser, webcam and microphone samples do not need to leave the computer. This makes it easier to justify classroom use under school data policies.
• Fast feedback loop: Small datasets usually train in seconds to minutes on an ordinary laptop. Learners can add examples, retrain and compare results several times within one lesson.
• Open-source libraries: The helper libraries and export snippets sit in a public GitHub repository under Apache 2.0. Developers can read exactly how models are loaded and adapt the code for their own projects.

Cons

Thin maintenance is the main worry, since Google labels it an experiment and the npm helper libraries were last updated in August 2021.

• Experimental status: The official repository says Google will try to support the tool but promises nothing. There is no service level, roadmap or guaranteed lifespan, so production systems should not depend on it.
• Classification only: Models can only sort inputs into the classes you define. There is no object detection with bounding boxes, no segmentation and no text or tabular data option.
• Little model control: The tool is built for simplicity, so you cannot choose or redesign the underlying model architecture. Accuracy therefore tends to level off on harder problems or classes that look alike.
• Uneven export coverage: Pose models export only to TensorFlow.js, and Keras export covers image projects only. Teams wanting mobile pose detection must convert models themselves or use another tool.
• Desktop browsers only: The training page asks visitors to use desktop Chrome or Safari and blocks unsupported devices. Tablets and phones are not suitable for building models, which can limit some classrooms.

Who is Using Teachable Machine?

Educators form the core audience, using Teachable Machine to show students how training data shapes a model’s behaviour.

• Teachers and students: STEM and computing classes use it to teach classification, bias and data quality. A lesson can move from collecting examples to testing a working model in one session.
• Artists and designers: Creative coders build interactive installations that react to gestures, poses or sounds. The p5.js export snippets make it easy to connect a model to a sketch.
• Makers and hobbyists: Hardware tinkerers run image models on Arduino boards or Coral devices for small sensing projects. The embedded guide covers wiring, data capture and deployment.
• Prototyping developers: Product teams test whether an idea is feasible before investing in a full machine learning pipeline. A quick Teachable Machine model can demonstrate a concept to stakeholders in an afternoon.

Pricing

Teachable Machine costs nothing for any feature, including training, testing and every export format (Google site and directory listings, October 2026).

Directories such as Futurepedia and There’s An AI For That also list it as fully free, so there is no conflict on cost. Google publishes no pricing page, so if that ever changes the tool itself will be the place to confirm current terms.

What Makes Teachable Machine Unique?

Training happens on your own machine rather than in Google’s cloud, which is rare for a free tool backed by a large provider. That design removes sign-up friction and keeps webcam and audio samples local by default, a practical advantage in schools.

The other differentiator is reach. Few no-code trainers export the same model to a web page, an Android app, a Coral Edge TPU and an Arduino microcontroller. Cloud services such as Google Vertex AI aim at professionals with budgets, so Teachable Machine fills a gap for learners and hobbyists.

Compatibilities and Integrations

TensorFlow is the backbone of every integration, with exports covering web, Python, Android and microcontroller targets.

• Browsers: Training works on desktop Chrome and Safari. Exported TensorFlow.js models run in any modern browser or in Node.js.
• Python and Keras: Image models download as a Keras .h5 file with a labels file and sample code for Pillow or OpenCV.
• Android and Coral: TensorFlow Lite exports cover Android for image and audio models, plus Coral Edge TPU for image models.
• Arduino: The embedded image model targets the Arduino Nano 33 BLE family through TensorFlow Lite for Microcontrollers.
• Google Drive: Projects, including training samples, can be saved to and reopened from your own Drive account.

Teachable Machine Tutorials

A first image model takes about ten minutes from opening the page to testing it live.

1. Open the site in desktop Chrome or Safari and choose Get Started, then Image Project.
2. Rename Class 1 and Class 2 to match what you want to recognise, adding more classes if needed.
3. Use the Webcam or Upload button to add at least a few dozen varied examples to each class.
4. Click Train Model and keep the tab open until training finishes.
5. Test the model in the Preview panel and add more examples where it makes mistakes.
6. Select Export Model, pick TensorFlow.js, TensorFlow Lite or TensorFlow, and copy the matching code snippet.

How We Rated It

Teachable Machine scores 4.0 overall, because simplicity and zero cost are balanced by narrow features and an uncertain maintenance outlook.

Summary

Teachable Machine is the best free starting point for anyone who wants to see machine learning work without writing code. It suits teachers, students, artists and makers who need image, sound or pose classifiers for lessons, installations and prototypes. The main limitation is depth, since it only classifies, offers few controls and is maintained as an experiment. For commercial products, move to a supported platform once the idea is proven. A practical tip: collect examples in varied lighting, angles and backgrounds, because data quality decides accuracy far more than any setting.