PySTEM Mini Bot

📋 Steps

Step 1: initial training data — choose up to 5 speech commands, then add 20 1-second audio samples for each and label them.

Step 2: train and refine — train the model and check the training accuracy, add more audio samples and check their predictions, then re-train.

Step 3: use the model with your robot in the IDE, or with the virtual robot on the Simulator page.

💡 Tip: over 100 audio samples are needed to train a decent model; a low training accuracy indicates audio samples are of bad quality or mis-labelled.

💡 Tip: be prepared when clicking Add Audio Sample button, you have 1 second to say the speech command.

💡 Tip: each model is a named session (bar below). Keep one per project; samples and the trained model are saved in this browser.

Model:

Model training is based on TensorFlow.js in browser. You audio samples are not saved to any server.

Creation Time

Audio Sample

Wave

Spectrogram

Speech Command

Prediction

Probability

Quantized Prediction

Quantized Probability