Show HN: I made an audio-to-MIDI model 60x more efficient (kobi.music)

🤖 AI Summary
KobiMusic has introduced Dettato, a groundbreaking audio-to-MIDI transcription model that demonstrates a remarkable 60x efficiency increase compared to its closest competitor, MuScriptor (1.3B parameters). With only 20 million parameters, Dettato outperforms both MuScriptor and Mirelo's closed-source model on several benchmarks while ensuring compliance with copyright laws by training solely on synthetic and Creative Commons material. This open-source model marks a significant advancement in the field, allowing users to leverage its outputs for commercial purposes without infringing on copyrights. The development of Dettato underscores a shift in approach within the AI/ML community from focusing predominantly on data volume to optimizing architectural design. By employing innovative techniques like the 'soundfont babies' method and treating audio transcription as a local problem, the model achieves improved accuracy while avoiding overfitting on synthetic data. KobiMusic's commitment to helping musicians rather than replacing them highlights the potential for responsible AI development to drive innovation in music technology, emphasizing that constraints can foster creativity and better results in model performance.
Loading comments...
loading comments...