Training a model for audio classification

Just like with images, a model can learn to distinguish sounds: a click versus a snap, a 'yes' versus a 'no', or one specific machine noise versus another.

Here is how it works:

1. Record many short examples of each sound (at least 20 per class, 1–2 seconds).
2. Include background noise as its own class — otherwise, silence will be classified as one of the sounds.
3. Train.
4. Test with new recordings the model has not heard before.

The model does not see the sound as you hear it — it sees the waveform, often converted into an image (spectrogram) that it processes like an image model.

Experiment with three classes (Teachable Machine has an audio mode):

Class	Number of examples
click	25
snap	25
background	25

Then deliberately test difficult cases:

• Click further away from the microphone.
• Click with music in the background.
• Have someone else click.

Most models perform significantly worse on all three. Why? All training examples were recorded close up, in a quiet room, by the same person. The model learned the recording situation, not just the sound.

Collect data again with variation and measure again. It is the same lesson as with images — but it feels clearer with audio.