Genre Classification of Saint Yared Qum Zema Using a Convolutional Neural Network

Music has always been more than just sound; for centuries, it has served as a vessel for history, faith, and identity. In the realm of modern science, researchers have developed ways to teach computers to listen, a field known as Music Information Retrieval. This discipline asks machines to do what human ears do naturally: recognize patterns, identify instruments, and sort songs into categories like jazz, rock, or classical. While computers have become quite good at analyzing the vast libraries of Western music, they often struggle with unique, traditional forms that rely on the human voice alone. These ancient traditions, passed down through generations without written scores or instruments, present a special challenge. Their sounds are subtle, their variations are fine, and their preservation is urgent, as the number of people who can teach them is shrinking.

In Ethiopia, the Ethiopian Orthodox Tewahedo Church guards a sacred musical tradition known as Zema. This is a form of spiritual chanting composed in the sixth century by Saint Yared, a scholar revered as the head professor of music for his time. Saint Yared organized these chants into three distinct styles: Geez, Ezil, and Araray. When these chants are performed without any musical instruments, using only the human voice, they are called Qum Zema. For centuries, the knowledge of how to distinguish between these three styles has lived in the minds of a few specialized scholars within church schools. However, as the number of these experts declines and the general public finds it difficult to tell the styles apart, there is a risk that this subtle art could be lost or misunderstood. The question facing researchers was whether modern technology could learn to hear these differences as clearly as a seasoned scholar.

A team of researchers from Debre Markos University set out to answer this question by building a computer model designed specifically to recognize the three genres of Saint Yared's Qum Zema. They did not try to teach the computer by manually listing the rules of the music, such as specific notes or rhythms. Instead, they used a type of artificial intelligence called a convolutional neural network, which is a system designed to learn by looking at images. To make this work, the researchers first had to translate the sound of the chants into something the computer could see. They took recordings of the vocal performances and broke them down into short, ten-second clips. Each of these clips was then transformed into a spectrogram, which is a visual map showing how the pitch and volume of the sound change over time. In this map, the horizontal axis represents time, the vertical axis represents pitch, and the brightness of the colors shows how loud the sound is at any given moment.

The researchers gathered a total of 1,555 of these ten-second audio clips from scholars at traditional church schools. They ensured that the recordings were clean by removing background noise and then converted every single clip into a spectrogram image. This collection of images became the training data for their new model, which they named the Saint Yared's Qum Zema Classifier. The system was fed these images and asked to learn the visual patterns that correspond to the Geez, Ezil, and Araray styles. To test if the system was truly learning, the researchers split their data, using 70 percent of the images to teach the model and holding back the remaining 30 percent to see how well it performed on music it had never seen before.

The results showed that the computer could indeed learn to distinguish between these ancient vocal styles. When the model was tested on the unseen audio clips, it correctly identified the genre 88 percent of the time. This was a significant achievement, especially considering that the three styles are very similar to one another and are sung without any instrumental help to guide the listener. The researchers compared their custom-built model against several other well-known computer systems designed for image recognition, such as ResNet, VGGNet, and AlexNet. While those other systems were larger and required more computing power, they did not perform as well on this specific task. The custom model achieved a higher accuracy rate than the others while being much smaller and faster to run.

The study also explored how different settings within the computer model affected its ability to learn. They found that a specific way of processing the information, known as the ReLU activation function, helped the model learn the patterns more effectively than other methods. The entire process was built on the idea that by turning sound into pictures, a computer could find the unique "fingerprint" of each musical style without needing a human to explain the rules. The researchers noted that the model's performance was limited by the fact that the dataset contained only vocal sounds with high similarities, making the task inherently difficult, but the 88 percent success rate demonstrated that deep learning could be a powerful tool for preserving intangible cultural treasures.

This work offers a practical path forward for protecting and teaching Saint Yared's musical heritage. By creating a tool that can automatically sort these chants, the researchers have provided a way to support the education of new scholars and ensure that the distinctions between Geez, Ezil, and Araray are not lost to time. The study suggests that while the task is complex, the combination of traditional knowledge and modern machine learning can bridge the gap between the past and the future. The researchers plan to continue this work by gathering more data and exploring ways to make the system even more accurate, but the initial success proves that a computer can be taught to listen to the sacred songs of Ethiopia with a new kind of understanding.