Google Teachable Machine

A free web tool for training image, sound and pose classification models in the browser and then exporting them as TensorFlow.js, TensorFlow or TensorFlow Lite.

Information last verified:

IllustrationIllustration of the Google Teachable Machine interface
An illustration, not a screenshot of the software: the "Image classification" project with three classes from the real digits-8x8 data before training. Open the projectThe icon is an illustration made by the Academy, not the official logo.

A free web tool from Google for training image, sound and pose classification models in the browser, with no account and no code. It is based on transfer learning: a pretrained model (MobileNet for images and poses, and Speech Command Recognizer for sounds) with your own classes added on top. Ideal for teaching the concepts of training, testing and bias to students and school pupils.

Tool card

NameGoogle Teachable Machine
DomainComputer vision and audio for teaching and quick prototypes
DeveloperGoogle (a project within A.I. Experiments)
Latest version we verifiedA web tool updated on the server (no version number for the user)
Operating systemsAny system with a modern browser, plus a camera or microphone when needed
LicenceCompletely free
Open source?Open-source core, some components closed
Licence and pricing detailsFree to use with no account. The Teachable Machine libraries and usage snippets are published as open source on GitHub (the teachablemachine-community repository), while the service itself is run by Google.
Technical requirementsA modern browser. Training happens inside the browser tab; the samples are not sent to any server unless you save the project to Google Drive. When you “upload the model”, only the model is published on Google's servers, without the samples.
Difficulty levelBeginner

Problem types

  • Image classification
  • Audio classification (one-second clips)
  • Pose classification

Supported algorithms and node name

  • Transfer learning on top of MobileNet (images and poses)
  • Speech Command Recognizer (sounds)
  • PoseNet for estimating body positions

Types of data

  • Images from the camera or files
  • Sounds from the microphone (one second per sample)
  • Body poses from the camera or images

Import and export

Download the samples of each class, save the project as a zip file in Drive or on the device, and export the model as TensorFlow.js (hosted with a link or downloaded) or convert it to TensorFlow and TensorFlow Lite for local use.

Tutorial guide

  1. Image Classification
  2. Audio Classification
  3. Pose Classification
  4. Collecting data
  5. Training the model
  6. Testing the model
  7. Exporting the model
  8. Educational applications

Image Classification

Choose Image Project, then create a class for each type (for example: pen, cup, book). Images are captured from the camera or uploaded from files.

Audio Classification

Choose Audio Project. First record the “background noise” (Background Noise), then your classes. Each sample lasts one second, and external mp3 files cannot currently be uploaded.

Pose Classification

Choose Pose Project. PoseNet estimates the joint positions, so the tool learns from the pose rather than the clothes or the background, but it is affected by where you stand in the frame.

Collecting data

Vary the samples: different lighting, different backgrounds, different angles and people. Similar numbers across classes reduce bias. And remember that the tool learns only from what it sees.

Training the model

Press Train Model. The Advanced options let you change the Epochs, Batch Size and Learning Rate, and the “Under the hood” button shows the accuracy and loss curves for each epoch, the accuracy of each class and the confusion matrix.

Testing the model

The Preview panel tests the model live with the camera or the microphone. Test with new examples that were not in the training: another background, another person, another microphone.

Exporting the model

Export Model: TensorFlow.js (upload and host with a link, or download), TensorFlow, or TensorFlow Lite for phone apps and small devices. Ready-made code snippets are available for each format.

Educational applications

Sorting recyclable waste from images, a game controlled by gestures, recognising animal sounds, and experiments about bias: train on your voice and test with a classmate's voice. Do not use it for medical diagnosis.

Related lessons

Related projects

Verification sources

Compare it with another tool