Redefining Music Genre Classification: The Yemeni Dataset Breakthrough

Automatic music genre classification often finds itself stuck in a Western-centric loop, ignoring the rich tapestries of global musical traditions. Enter the Yemeni Music Information Retrieval (YMIR) dataset, a notable departure from the norm. This dataset isn't only a nod to cultural diversity but also a technical milestone with 1,475 meticulously curated audio clips across five traditional Yemeni genres: Sanaani, Hadhrami, Lahji, Tihami, and Adeni.

A Dataset with Depth

The YMIR dataset didn't just slap labels on tracks. Instead, the classification was handled by a cadre of five Yemeni music experts, adhering to a structured protocol. This approach resulted in a Fleiss kappa of 0.85, showcasing remarkable inter-annotator agreement. Such rigor provides a level of authenticity and reliability that many datasets merely aspire to.

Meet the Yemeni Music Classification Model

At the heart of this venture is the Yemeni Music Classification Model (YMCM). Built on a convolutional neural network (CNN), YMCM doesn't just tick boxes, it sets a new standard. The model was put through its paces with a consistent preprocessing pipeline, engaging in 30 different experiments across six groups and five architectures.

Feature representations like Mel-spectrograms, Chroma, FilterBank, and MFCCs with varied coefficients were evaluated. In this rigorous benchmarking, YMCM didn't just hold its own. It dominated, achieving a staggering 98.8% accuracy with Mel-spectrogram features. If that doesn't make you rethink the potential of culturally specific datasets, what will?

The Bigger Picture

Why does this matter? Because slapping a model on a GPU rental isn't a convergence thesis. It's about diversifying the datasets that feed our models. The YMIR dataset provides new insights into the relationship between feature representation and model capacity, a connection often overlooked in homogeneous datasets.

But the question remains: How will this influence future models? It's clear the intersection is real. Ninety percent of the projects aren't. However, those that succeed, like this one, redefine the boundaries of music information retrieval. To skeptics, I say, show me the inference costs. Then we'll talk.