What they built.
Steve Saling's collaboration begins with a specific person's available movements and sounds. Google's 2019 account describes facial gestures used to cheer during a game and non-speech sounds used for smart-home commands. The useful output is an action that the person can intentionally trigger.
This is personalized classification: examples teach a model to distinguish signals chosen for a particular use. Google's later Teachable Machine 2.0 release makes a related process available in a browser. The published demonstration establishes a concrete prototype, while its full custom application and individual performance remain undocumented.
Google identifies Saling and describes concrete demonstrated actions. The public tool provides a related train/test/export workflow; Saling's complete custom configuration is not supplied.
Read the original accessibility demonstrationHow the tools work together
In the public tool, a person gathers examples, trains a model and tests it against new input. TensorFlow.js handles browser-based learning. A recognized category can then be connected to an action by an application; training the classifier alone does not create the complete device-control workflow.
Workflow at a glance
Gather personal examples
Record examples of the chosen image, sound or pose classes with a webcam or microphone.
Train and test the recognition
Teachable Machine trains in the browser; test new examples and revise the training examples where recognition fails.
Connect a recognized class to an action
The demonstrated prototype connects personalized recognition to sounds or device commands. An exported model requires application code for that connection.
How you could adapt the idea
Begin with a harmless feedback action, such as playing a sound, and involve the intended user in choosing comfortable, repeatable signals. Try examples in the actual setting and inspect mistaken activations. This proposed learning exercise does not replace a person's established communication system or demonstrate clinical effectiveness.
Effort, cost and limitations
- Effort
- A camera or microphone and browser suffice for the recognition exercise. Reproducing the external action needs additional integration work.
- Cost
- The sources do not disclose Saling's project cost or a complete reproduction budget.
This is a personalized research demonstration, not evidence of clinical benefit or a universally reliable control system.
The 2.0 documentation describes the later public tool; it does not establish an identical model or interface in the earlier demonstration.
Go to the original work.
Original Project Euphonia demonstration and named collaboratorSteve Saling project in Start With OneTeachable Machine 2.0 train/test/export documentationPublic Teachable Machine interfaceSource-backed editorial account; PrisAIm did not run or independently benchmark this implementation.