Write a word or a sentence in the box with your mouse, and the project will try to read it as you write. USE THIS ⬇ IT WONT WORK ON NORMAL SCRATCH https://turbowarp.org/1392124543 How to write: 1. Print each letter separately. Leave a small gap between letters and a bigger gap between words. 2. Keep writing. The result updates live while your pen is down. 3. Click "Clear" or press C to start over. What you will see: - A blue box around each letter the project found. If two letters share a box, they were written too close together. - Each letter as the neural network sees it, with its best guess underneath. - The recognized text, plus other possible readings of the word you are writing. Tips: - Cursive does not work yet. Print your letters. - Single letters and digits work too. - Press T to run a self-test that checks the neural network against its original version. - The project does a lot of math every frame. If it feels slow, try opening it in TurboWarp.
This project reads handwriting with a real neural network running entirely inside Scratch. How it works: 1. Your strokes are grouped into letters. Small marks, like the dot on an "i", are attached to the nearest letter. 2. Each letter is scaled and centered into a 28x28 image, the same format the network was trained on. 3. A convolutional neural network with 109,246 weights classifies each letter as one of 62 characters (0-9, A-Z, a-z). The weights are stored in a list, and every multiplication happens in Scratch blocks. 4. Each word is compared against a list of about 9,300 common English words, which helps when letters look alike (for example "he110" becomes "hello"). If nothing in the list fits, it shows the letters it read. The network was trained in Python with PyTorch. Credits: - EMNIST dataset by Gregory Cohen, Saeed Afshar, Jonathan Tapson and André van Schaik (2017), based on NIST Special Database 19 - Word list from google-10000-english by first20hours (github.com/first20hours/google-10000-english) - PyTorch, used to train the network Known limits: letters that look the same when handwritten, like c and C or o and 0, can be mixed up, and cursive writing is not supported. #AI #NeuralNetwork
