Flute Tone Quality Classification: a Machine-Learning-Based Evaluation Tool

Sane N, Abel J. Flute Tone Quality Classification: a Machine-Learning-Based Evaluation Tool. In: 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio. Audio Engineering Society; 2025. Paper 6. Available from: https://aes.org/publications/elibrary-page/?id=22995

@inproceedings{Sane2025_22995,
  author = {Sane, Nikita and Abel, Jonathan},
  title = {{Flute Tone Quality Classification: a Machine-Learning-Based Evaluation Tool}},
  booktitle = {2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio},
  note = {Paper 6},
  year = {2025},
  month = sep,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=22995}
}

TY  - CPAPER
TI  - Flute Tone Quality Classification: a Machine-Learning-Based Evaluation Tool
AU  - Sane, Nikita
AU  - Abel, Jonathan
T2  - 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio
M1  - Paper 6
PY  - 2025
DA  - 2025/09/02
UR  - https://aes.org/publications/elibrary-page/?id=22995
PB  - Audio Engineering Society
LA  - en
AB  - Music classification models have been used to accurately place audio recordings into different classes based on genre, instrument, and era, among other features. In this work, we introduce a machine learning algorithm that classifies flute recordings based on their tone quality. Flute tone quality is subjective, and its underlying parameters currently lack clear definitions in terms of objective audio signal processing features. Here we consider two quality descriptors, "focus" and "brightness," and present a largely "black box" algorithm to categorize recorded flute notes into quality descriptor classes. The model was built by first gathering and labeling a set of roughly 5000 high-quality flute note recordings derived from the Freesound database according to three levels each of focus and brightness. Then a pre-processing script that performs spectral analysis was developed. Finally, a convolutional neural network was trained, and its hyper-parameters, including the learning rate, batch size, and dropout rates, adjusted. An overall classification rate of 77% was achieved, with brightness and focus being correctly classified at rates of 83% and 86%, respectively.
ER  -