Thread by @gusthema

To start with Audio classification, the easiest way is with Google's Teachable Machine: teachablemachine.withgoogle.com

You won't need any expertise, just add some audio samples and you'll have a model

Let's take a quick look on the basics and some internals

#teachablemachine

1/8🧵

You can gather the data right in the the tool:

There are some great tips on how to gather better data here 🗣️🎙️:
teachablemachine.withgoogle.com/faq#Tweaking-y…

2/8🧵

Some tips are 💡:

-> To improve the quality of your audio data, try recording it with different microphones, in different rooms, changing the proximity to the mic.
-> The model will learn from 1-second samples so it won't be able to recognize longer events.

3/8🧵

After you have some data to train a model, you can define the number of epochs and start training your model

The training runs completely in the Browser using TensorFlow.js

Your data doesn't go to any server unless you choose to save it to your Google Drive

#MadeWithTFJS
4/8🧵

With the model trained you can Export it and/or just play with it on the browser

Export your model:

Here you can define an Overlap Factor, which means how frequently the last second of audio is classified by the model

5/8🧵

You can export the model to run on the Browser with TensorFlow.js or on a mobile device with TensorFlow Lite.

They even give you all the source code you'll need.

For the full code to run on Android, you can check out this repository: github.com/tensorflow/exa…

6/8🧵

For Audio, the model is built using the Speech Command Recognizer sample (github.com/tensorflow/tfj…)

Would you like a more detailed view on this sample specifically?

7/8🧵

All the source code for Teachable Machine is available here: github.com/googlecreative…

Teachable Machine is a great start!

But what would you do if you need a more robust base model? or have a big dataset?

I'll answer this questions later, stay tuned! 👀

8/8🧵