Small data AI replicates Persian calligraphy sounds

Researchers developed a small-data generative AI model to recreate the unique sounds of Persian calligraphy.

The team used a bespoke audio dataset to train two specific models, RAVE and StyleGAN2. These models were adapted to the sonic domain through audio synthesis and spectrogram generation to capture the fine-grained timbral and gestural nuances of sarir, the distinctive sound produced by a reed pen during the creation of Persian calligraphy.

The results show that these small-data models perform better than state-of-the-art text-to-audio models when dealing with highly situated cultural contexts. Because large-scale AI often reproduces cultural biases or fails to engage with non-Western traditions, this approach offers a way to preserve and generate culturally specific artistic materials. The research demonstrates that focusing on smaller, artist-curated datasets can unlock expressive possibilities that broader models cannot reach. Future work may extend this method to other cultural materials and artistic domains.