Automatic Identification of Maxaatiri and Maay Somali Dialects from Speech Using Mel-Spectrogram-Based Convolutional Neural Networks

Imagine you have a giant library of audio recordings, but instead of books, they are voices speaking different versions of the Somali language. The goal of this study was to build a digital "ear" that can listen to a short clip of speech and instantly tell you: "Is this the Maxaatiri dialect (often called Standard Somali) or the Maay dialect?"

Here is how the researcher, Mohamed Mohamud Ali, built this system, explained in simple terms:

1. The Problem: A Missing Translator

Somali is spoken by millions, but it's not just one single voice. It has major dialects, like Maxaatiri and Maay. Think of them like the difference between a British accent and an American accent, but with bigger differences in how words are built and pronounced.

Currently, most computer programs that understand speech are trained on "Standard" languages. They often get confused when they hear a different dialect, much like a person who only knows how to read formal textbooks might struggle to understand a friend's local slang. This study wanted to see if a computer could learn to tell these two Somali dialects apart just by listening.

2. The Recipe: How the Data Was Gathered

Since there wasn't a pre-made "Somali Dialect Dataset" available in a box, the researcher had to cook up their own ingredients.

• The Source: They went to YouTube and downloaded public broadcast videos.
  • For Maxaatiri, they used clips from Somali National Television.
  • For Maay, they used clips from Arlaadi TV.
• The Cutting Board: They chopped these long videos into tiny, 10-second slices of pure speech.
• The Cleanup: They removed silence and background noise, turning everything into a clean, standard audio format (like converting all files to the same MP3 quality).

In the end, they had a "training meal" of 180 audio clips: 90 from each dialect. It was a small, balanced meal, but enough to start tasting.

3. The Magic Glasses: Turning Sound into Pictures

Computers are great at looking at pictures, but they are bad at listening to raw sound waves. To fix this, the researcher used a trick called a Mel-Spectrogram.

Imagine taking a sound wave and running it through a prism. Instead of seeing a rainbow of colors, you get a heat map or a sonic fingerprint.

• The X-axis is time (how long the sound lasts).
• The Y-axis is pitch (how high or low the sound is).
• The colors show how loud different pitches are.

Now, instead of a sound file, the computer has a little picture of the sound. This makes it much easier for the computer to "see" the differences between the two dialects.

4. The Detective: The Neural Network

The researcher built a Convolutional Neural Network (CNN). Think of this as a digital detective trained to look at those "sound pictures."

• Training: The detective was shown 180 pictures. Some were labeled "Maxaatiri" and some "Maay."
• Learning: The detective looked for patterns. Maybe Maxaatiri sounds have more "red" in the high-pitch area, while Maay sounds have more "blue" in the low-pitch area.
• The Test: After studying, the detective was given a new set of pictures it had never seen before (the test set) and asked to guess the dialect.

5. The Results: How Good Was the Detective?

The detective did a pretty good job, but it wasn't perfect.

• Accuracy: It got the answer right about 81.5% of the time.
• The Mistakes: Out of the test clips, it correctly identified 12 out of 14 Maxaatiri clips and 10 out of 13 Maay clips.
• The Confusion: Sometimes it mixed them up. It thought a few Maxaatiri clips were Maay, and a few Maay clips were Maxaatiri.

6. The Big "But": Why We Should Be Careful

The paper is very honest about its limitations. It's important to understand why the detective might be good at this specific test but not ready for the real world yet.

• The "TV Station" Bias: This is the biggest catch. The researcher used one TV station for Maxaatiri and a different TV station for Maay.

  • Analogy: Imagine you are trying to teach a child to tell the difference between "Apples" and "Oranges." But you only show them Red Delicious apples from a specific grocery store and Green Oranges from a different grocery store.
  • The child might not be learning the difference between the fruit; they might just be learning the difference between the grocery store logos or the lighting in the store.
  • Similarly, the computer might have learned to recognize the sound quality of Somali National TV vs. Arlaadi TV, rather than the actual dialect. If you played a Maay speaker on Somali National TV, the computer might get confused.

• Small Sample Size: The dataset was tiny (only 180 clips). It's like trying to learn a whole language by reading just two short stories.

• Missing Details: The researcher didn't know who the speakers were (their age, gender, or region). If the Maxaatiri speakers were all men and the Maay speakers were all women, the computer might just be guessing "Male vs. Female" instead of "Dialect A vs. Dialect B."

The Bottom Line

This paper is a first step. It proves that it is possible to use deep learning to tell these two Somali dialects apart using sound pictures. It's like building a prototype car engine that runs on a test track.

However, the paper does not claim this engine is ready for the highway yet. Because the data came from limited TV sources and was small, the results are just a "preliminary baseline." To make this technology useful for real people, we would need a much bigger library of voices from many different speakers, recorded in many different places, so the computer learns the dialect and not just the TV station.