Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Teachable Machine is a free, browser-based tool from Google Creative Lab for training simple image, sound, and pose classifiers without writing code. You provide labeled examples, train a model, test it, then download or host it for use in a website, app, or supported maker project. It is useful for learning, creative experiments, and prototypes—not a general-purpose AI or a dependable system for high-stakes decisions.
What Teachable Machine is—and what it is not
Teachable Machine makes one part of supervised machine learning approachable: teaching a model to sort new inputs into categories using examples you label. For instance, you could show it examples of ripe and unripe fruit, then test whether it can classify a new image.
The model does not understand what “ripe” means. It learns statistical patterns in the supplied examples and predicts which class a new input most resembles. Those patterns can be useful, but they can also be accidental: the model might distinguish the backgrounds or lighting in your examples rather than the objects you intended. Confidence scores are not proof of correctness or calibrated probabilities.
Free tools Windows power users keep installed
One-click scans. No signup required.
The tool is not a chatbot, speech-to-text service, universal object recognizer, or substitute for a full computer-vision platform. It is best treated as an educational and prototyping interface. Google Creative Lab describes the project as an experiment, and its community code repository says it is not an official Google product. Teachable Machine is therefore not a conventional Google Cloud enterprise service.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
Is Teachable Machine still available?
Yes. The current interface is at teachablemachine.withgoogle.com/train, where the available project types include image, sound, and pose. Use this current interface to create and export a model. The older 2017 Teachable Machine experiment remains available as a separate legacy experience; it is historically useful, but it is not the current project workflow.
The original experiment introduced a simple way to teach a computer from examples. Google’s announcement of Teachable Machine 2.0 described a broader no-code workflow for training models and exporting them to websites, apps, and physical projects, with support for images, sounds, and poses. Interface labels and export options can change, so check the live page for the controls available to your project.
What can it classify?
Images
An image project can use a webcam or image files and distinguish user-defined categories. Example projects include recognizing hand gestures, sorting a few kinds of recyclable material, or triggering a simple game action. An image classifier can mistake a background, camera angle, or lighting condition for the category itself. Include varied examples and test in the place where the model will actually be used.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Rank #2
Sounds
A sound project classifies short audio examples, such as a clap, snap, whistle, doorbell, or silence. The current interface describes examples around one second long. Do not assume every audio file type is accepted: supported file inputs may vary, so check the live project interface. This is short-sound classification, not robust speech recognition or transcription. Noise, echoes, microphone differences, and recording distance can all affect results.
Poses
A pose project can distinguish body configurations such as arms raised or lowered, standing or sitting, or a head tilted left or right. It can support simple hands-free controls, games, or demonstrations. Camera framing, lighting, clothing, occlusion, and distance matter; a pose classifier is not a comprehensive system for understanding human actions, and performance with multiple people depends on the training setup.
How to make and test a model
For a first project, use a modern desktop browser. Image and pose projects need camera permission; sound projects need microphone permission. A clear class definition and an uncluttered, reasonably well-lit environment make it easier to spot errors. Obtain permission before using other people’s images, voices, or recordings.
- Open the training page. Go to the current training interface, rather than the legacy
/v1/experiment. - Choose a project type. Select Image Project, Audio Project, or Pose Project, according to the input you want to classify. These are current interface labels, not a promise that every option or export format will remain unchanged.
- Define the classes. Create a class for each category, such as
Ripe,Unripe, andBackground, orClap,Snap, andSilence. A neutral or “none of the above” class gives the model examples of inputs that should not trigger a target category. Without it, an unfamiliar input may still be assigned to one of the target labels. - Gather varied examples. For images, vary object orientation, distance, background, lighting, and camera position. For audio, vary volume and distance, and capture representative room noise. For poses, include the neutral position and modest variations in position, distance, clothing, and camera angle. The legacy experiment suggests at least 30 images per image class as a teaching tip, not a universal threshold for accuracy.
- Train the model. Choose Train Model and wait for the preview to become available. The current site describes training as browser-based, and Google’s announcement says training runs locally on the user’s computer. Keep the page open; device performance, available memory, permissions, and a suspended tab can affect the process.
- Test on examples it has not seen. Try new images, sounds, or poses—not just the examples used to train it. Change the lighting or background, use the actual deployment microphone, or ask another person to try it. Note which classes are confused. A confident wrong answer is still wrong.
- Export when the behavior is good enough for the intended experiment. Use Export Model and choose an available format. The site offers model downloads and online hosting; options differ by project type and may change. Test the exported model in its actual runtime before relying on it.
A practical example: ripe versus unripe fruit
Suppose students want to make a camera-triggered fruit-sorting demonstration. Create classes for Ripe, Unripe, and Other/background. Gather examples of both fruit categories under several backgrounds and lighting conditions. Include different fruit orientations and distances; do not put all ripe fruit on one table and all unripe fruit on another, or the model may learn the tables.
Recommended Free Tools
After training, test with fruit and backgrounds not used during training, including objects that are neither ripe nor unripe fruit. If those objects receive a confident fruit label, collect better examples for the neutral class and retest. The result is a useful classroom demonstration of classification and data quality, not a reliable food-safety judgment or an industrial sorting system.
Why a model fails, and what to change
- It recognizes the background instead of the object: The classes may have been photographed in different places. Recollect examples with mixed backgrounds, vary lighting and distance, add neutral examples, and test in a separate location.
- It works only for the person who trained it: It may have learned a person’s hands, clothes, voice, or posture. Where appropriate, include examples from multiple people and hold out at least one person for testing.
- Every input gets a target label: Add a neutral class containing unrelated objects, silence, empty scenes, or resting poses that resemble real non-trigger inputs.
- Examples are too similar: Repeating nearly identical frames may not help. Broaden the examples across the conditions the model will encounter.
- Sound recognition changes in another room: The model may have learned echo, background noise, microphone characteristics, or volume. Record target and neutral sounds under representative conditions, then test with the deployment microphone.
- Pose recognition breaks at a different distance: The model may associate a pose with a particular camera framing. Include examples at relevant distances and ensure the camera captures the body landmarks needed for the task.
- The preview looks good but the app fails: Preview tests may be too similar to training data. Keep held-out examples and test with the actual camera, microphone, device, and runtime.
For browser or device problems, first check camera or microphone permission, then reload the project and close memory-heavy tabs. Try a current desktop browser if a phone or older device struggles. A smaller project with fewer or shorter examples may train more comfortably. For an offline deployment, download the model and confirm that its selected runtime and any other required assets are available locally.
Rank #4
Privacy: what local training means
The official site says the tool can run entirely on-device, without webcam or microphone data leaving the computer, and Google’s announcement says training examples remain on the device unless the user chooses to save the project to Google Drive. That describes the local-training mode; it is not an unconditional guarantee about every browser, extension, operating-system service, or later sharing choice.
Saving a project, uploading files, sharing a model, or using hosted assets can create separate data exposures. Before using sensitive biometric, medical, workplace, or children’s data, review the current FAQ, privacy information, and terms, and choose examples you are permitted to use. Avoid treating “training happens locally” as equivalent to “there is no privacy risk.”
Exporting and using a model
Export separates no-code training from application development. Training a basic model does not require programming, but putting it into a useful website, mobile app, or hardware project usually requires some coding or a suitable development environment.
Best Value
- Hosted model: Convenient for a quick web prototype because the model is accessed online. It depends on network access and the continued availability of the hosted assets.
- Downloaded model: Gives you local files for a project that needs more control, reproducibility, or reduced dependence on hosted assets. It does not by itself make an application work offline; the app must include and load the required files locally.
- TensorFlow.js: The common route for JavaScript and browser projects. The community repository provides helper libraries, snippets, and examples for model integration.
- Embedded workflows: The site lists compatibility with tools and hardware including p5.js, Node.js, Coral, and Arduino. Specific exports vary. The repository’s Arduino Nano 33 BLE/Nano 33 BLE Sense and OV7670 example uses a more advanced TensorFlow Lite for Microcontrollers workflow; it is not a plug-and-play recipe for every Arduino board.
Check the export panel and integration documentation for the supported format for your particular project. A model that works in the preview may still fail in deployment because of different camera resolution, mobile-browser constraints, network dependency, memory limits, or an unsupported runtime.
Teachable Machine for classrooms and creative projects
Its strongest educational use is making the relationship between examples and predictions visible. Students can change a dataset, retrain, and see how a classifier changes—an accessible way to discuss training data, testing, class design, and bias. Ask who or what is represented in the examples, which conditions were sampled, which were left out, and whether the model works for people or places not included during training. Google’s educational materials include lessons on AI ethics and bias.
For artists and makers, a small classifier can become an input for a browser sketch, interactive artwork, game, or physical prototype. The site lists JavaScript, p5.js, Glitch, Node.js, Coral, and Arduino among compatible environments. Hardware projects may require additional setup, code, and compatible components; no hardware purchase is necessary for a browser-based experiment.
How it compares with other tools
- Machine Learning for Kids: A better fit when the priority is guided classroom activities and lessons; it is less focused on a direct model-export workflow.
- MIT App Inventor: Useful when the goal is a block-based mobile app. Research has documented ways to deploy Teachable Machine image models in App Inventor projects, but this is an app-building path rather than a replacement for the training tool.
- Wekinator: A creative-machine-learning alternative for artists and interactive systems; it is also cited as an inspiration for the original Teachable Machine experiment.
- TensorFlow.js directly: Better for developers needing control over preprocessing, model architecture, evaluation, training, and application logic. It requires substantially more technical knowledge.
- TensorFlow Lite or Lite Micro directly: Better suited to embedded deployment and optimization, but requires managing conversion, hardware constraints, memory, and device-specific toolchains.
- Cloud machine-learning platforms: More appropriate when a production system needs managed deployment, monitoring, access controls, data pipelines, or scalable inference. They are more complex and can involve separate infrastructure and data-governance considerations.
Is Teachable Machine suitable for production?
Usually not by itself for production systems that need dependable performance across devices, people, locations, and conditions. It does not provide a guarantee of accuracy, an auditable evaluation, or the full set of controls expected for large-scale data management, monitoring, retraining, and governance. Do not use an unvalidated Teachable Machine model to make safety-critical, medical, legal, security, or industrial-control decisions.
It can still be a useful first prototype. Before using any model beyond a demo, evaluate it on a representative held-out dataset and in the deployment environment, decide how to handle uncertain or unexpected inputs, and verify hosting, privacy, runtime, and maintenance requirements. If those needs exceed a small prototype, move to a platform and workflow designed to meet them.
Cost and requirements
Teachable Machine is presented as a free web tool, and its official site says online model hosting is available for free. No paid Teachable Machine subscription is identified in the official pages cited here. A browser prototype needs no required hardware purchase, while an application, hosting service, or physical-computing setup may have separate costs. The tool’s free access does not imply guaranteed long-term hosting, uptime, or API stability.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
