View on GitHub

music-instrument-classifier

This is a simple classifier that is able to detect single-note sounds of various musical instruments. Currently supported types are cello, clarinet, flute, violin and piano.

Music Instrument Classifier

This is a simple classifier that is able to detect single-note sounds of various musical instruments.

Currently supported types are cello, clarinet, flute, violin and piano. The audios are supposed to be single-note sounds within the 4th octave.

Created by Ivy Zheng, Weiyuan Deng and Yuchen Rao from Northwestern University.

Dataset

The dataset used to train this classifier was collected from London Philharmonic Orchestra Dataset (http://www.philharmonia.co.uk/explore/sound_samples). Each audio file records one note from one of the five instruments, and has a length from 0.25 seconds to 6 seconds. In total, over 600 audio pieces were used to train the classifier, and the distribution of each class was roughly the same.

For copyright reasons, the original data is not included in this repository.

Dependencies

Python 2.7
NumPy
Scikit-Learn (sklearn)
Librosa
glob

Implementation

Preprocessing

The music pieces have their leading and ending silence trimmed. The threshold of trimming is 0.001 - if the intensity of the sound in the frame is below 0.1% of the highest sound intensity in the audio file, then the frame is trimmed out.

Feature Extraction

The Mel Frequency Cepstral Coefficents (MFCCs) of each music piece was extracted using Librosa. For each audio file, its MFCCs are averaged to produce the final, length-20 feature vector.

Classification

An SVM classifier is trained from the feature vectors to determine the instrument it belongs to. The SVM classifier worked in a one-vs-rest fashion, in which it trained 5 classifiers for each intrument class against all audios that are not in that class. The kernel of the SVM classifiers is linear.

Performance

Accuracy

With the given dataset and under 10-fold cross-validation, we achived a 95% accuracy on the instrument label prediction.

Time Efficiency

The most time-consuming part of the system is the MFCC feature extraction. However, it still needs less than a second to process a one-second audio, which makes it feasible to do real-time instrument detection.

Try it Out

In addition to the code for training, this repository also includes a pre-trained model that you can play with. Try to run the following code in your terminal:

python demo.py

After that, type in the audio file path that you want to classify. It will output the musical instrument that it recognized in the audio.

We also have a short video of this demo here: https://youtu.be/NSV8a49cqXU

Greg Sell, Gautham J. Mysore, and Song Hui Chon. Musical Instrument Detection. Thesis. Center for Computer Research in Music and Acoustics, 2006. N.p.: n.p., n.d. Web. 17 Feb. 2017.

Christian Simmermacher, Da Deng, and Stephen Cranefield. Feature Analysis and Classification of Classical Musical Instruments: An Empirical Study. Thesis. Department of Information Science, University of Otago, 2006. N.p.: n.p., n.d. Web. 17 Feb. 2017.

Sabin, Manuel. Musical Instrument Recognition With Neural Networks. Thesis. 2013. N.p.: n.p., n.d. Web. 17 Feb. 2017.

Agostini, Giulio, Maurizio Longari, and Emanuele Pollastri. Content-Based Classification of Musical Instrument Timbres. Thesis. Laboratorio Di Informatica Musicale - L.I.M, n.d. N.p.: n.p., n.d. Web. 17 Feb. 2017.