This page contains audio samples that demonstrate the biases found in our paper "GlobalDISCO: Evaluating Global Inequalities in AI-Generated Music". We present four challenging regional genres using examples from four state-of-the-art music generation models (Suno, Mureka, Riffusion, Udio), which performed worse on niche musical styles and generated songs that are perceptually closer to tracks from other (often more mainstream) genres.

Each row consists of the following components:

• Prompt: The musical style prompt from the GlobalDISCO dataset that was used as input to the model.
• Reference: A YouTube track from the artist whose musical style the prompt describes. This serves as a ground truth example.
• Generated: The audio output from the AI model based on the prompt.
• Nearest Neighbor: The closest YouTube track from a mainstream genre (pop or rock). The distance between audio tracks was calculated and aggregated across multiple embedding spaces.