Musical instrument classifier for early childhood percussion instruments

Area of Science:

• Music Information Retrieval
• Human-Computer Interaction
• Educational Technology

Background:

• Identifying non-pitched percussion instruments presents challenges due to sound variations.
• Existing musical instrument classification often overlooks these instruments, crucial for early childhood music education.

Purpose of the Study:

• To develop and evaluate a robust musical instrument classifier for tambourines, maracas, and castanets.
• To create a large-scale dataset of non-pitched percussion instrument sounds.
• To assess the model's performance in both lab and real-world home settings, including mixed reality applications.

Main Methods:

• Generated a diverse dataset of over 369,000 in-lab and 35,361 home-recorded non-pitched percussion instrument samples.
• Conducted sensitivity analyses for feature selection, windowing, and model optimization.
• Utilized a Light Gradient Boosting Machine (LGBM) model with mel-frequency cepstral coefficients (MFCCs) and signal entropy.

Main Results:

• The LGBM model achieved over 84% accuracy in lab settings and over 73% accuracy in home settings.
• Optimal performance was achieved using a 93 ms window with 12 MFCCs and signal entropy.
• A low feature space was found to be sufficient for recognizing non-pitched instruments.

Conclusions:

• The developed classifier effectively identifies non-pitched percussion instruments, particularly for early childhood music education.
• The large, novel dataset and efficient model contribute to advancing musical instrument recognition.
• Real-world deployment in a mixed reality application highlights potential for inclusive educational tools.