Simply making speech recognition data in the Machine Learning of "Teachable Machine" made by Google ~ using TensorFlow Technology in "Scratch"

After the tool is opened, register the topmost background noise first. Let you learn the sound of doing nothing as a pattern. As shown in the following figure, only 20 seconds of noise are recorded, each producing 20 voice samples per second.

Next, register the pattern you want to learn. Next time, the work produced in "Scratch" wants to use a speech recognition model to determine which sound is produced by "pineapple" or "apple". Don't let the pronunciation of "pineapple" and "apple" be learned as a pattern.

First of all, follow the steps shown in the following picture to start with the learning sample of "pineapple". Each time you record, sound as "pineapple", and then create a "pineapple" voice sample.

This tool is a mechanism that allows you to learn a lot of voice samples in one second. Each model needs more than 8 samples, and more samples are the learning data that identify all kinds of "pineapple".

The recording takes two seconds at a time, but it is automatically split into two samples one second at a time. The goal this time is to identify the pronunciation of a particular word, so delete samples that do not contain pronunciation and add only some samples such as "panau" and "pull". If you do not do so, the accuracy of recognition will be reduced.

Similarly, we will create an "Apple" learning sample. After the learning samples of "background noise", "pineapple" and "apple" are complete, press the [training model] to let you carry out machine learning.