How to Classify Blueberries as "Crunchy", "Juicy" or "Soft" using Acoustic Signal Processing and Machine Learning?

Question

I'm working on a project to classify blueberries based on their texture—specifically, whether they are soft, juicy, or crunchy—using the sounds they produce when crushed.
I have about 1100 audio samples, and I've generated spectrograms for each sample. Unfortunately, I don't have labeled data, so I can't directly apply supervised machine learning techniques. Instead, I'm looking for effective ways to differentiate between these three categories based on the spectrograms. I've attached examples of spectrograms for what I believe might be soft, juicy, and crunchy blueberries. However, since the data isn't labeled, I'm unsure if these assumptions are correct.
This is my Hypothesis:

Crunchy Berries: When crushed, they produce separate, distinct peaks in the audio signal. These peaks are spaced out over time, indicating that the berry is breaking apart in a crisp, segmented manner.

Juicy Berries: When crushed, they generate continuous peaks in the audio signal. These peaks are more closely packed together and sustained, indicating a burst of juice and flesh, with less resistance, creating a smoother sound.

Soft Berries: These produce very few and small peaks. The sound is faint and less defined, indicating that the berry crushes easily with little resistance, creating minimal disruption in the audio signal.

Answer 1

You could try doing semi-supervised learning. This is a machine learning technique that uses a small amount of labeled data and a large amount of unlabeled data.

In particular, you could attempt to label by hand for a small amount of data instances whether the blueberry is crunch, juicy, or soft. Then, you use semi-supervised learning on that small labeled dataset and the rest of the other unlabeled dataset.

Yes, as you stated, a problem with unsupervised learning is that you do not know whether your hypothesis is correct. However, there are unsupervised learning techniques, such as clustering, that you could do.

I hope this is helpful!