ANN model identifies 20 instruments using only MFCC audio features
Researchers train a neural net on the London Philharmonic dataset, achieving state-of-the-art accuracy…
In a paper published on arXiv (May 2021, updated May 2026), researchers from India propose a deep neural network for musical instrument recognition that relies solely on MFCCs—a compact representation of the short-term power spectrum of sound. The model is an artificial neural network (ANN) trained on the London Philharmonic Orchestra dataset, which includes 20 instrument classes spanning four families: woodwinds, brass, percussion, and strings. By extracting 13 MFCC coefficients from each audio sample and feeding them into a fully connected network, the system learns to distinguish instruments without requiring complex spectrograms or pretrained embeddings.
The experimental results show that the proposed method achieves state-of-the-art accuracy on this dataset, outperforming prior work that used richer feature sets or deeper architectures. The approach is computationally efficient and generalizable, suggesting that MFCC-based ANNs can serve as a strong baseline for automatic music classification tasks. This work has implications for music information retrieval, automatic transcription, and educational tools that need to identify instruments in real time with minimal compute resources.
- Uses only MFCC features (13 coefficients) rather than raw audio or spectrograms.
- Trained and tested on the full London Philharmonic Orchestra dataset with 20 instrument classes.
- Achieves state-of-the-art accuracy with a simple fully connected ANN model.
Why It Matters
Lightweight MFCC-based models can enable real-time instrument recognition on edge devices, powering music education and production tools.