Detecting instruments in classical music: A view on timbre, musical features, and classification

Abstract

This thesis describes research on recognition of classical instruments from isolated tones and musical passages. The intention is to evaluate the robustness of the features and the efficiency of the data analysis methods to classify instruments in a dynamic environment.

The information extracted from the samples is based on relevant works in instrument detection and timbre studies, and on knowledge about instrument characteristics. For a better understanding of the perceptual derived features, the human auditory system is briefly reviewed and some perceptual phenomena are mentioned. Different feature extraction methods are chosen without applying sound source separation techniques.

These methods are based on auditory models and filters and are widely used for simulation of the human perception and speech detection. Furthermore, a standardised set of features from the MPEG-7 Instrument timbre description scheme is calculated from the harmonic spectrum of a sound. It is especially designed to capture the timbre of instrument tones.

Two experiments are discussed in this thesis. The first experiment is used to evaluate features and classification methods on a dataset of single tone samples from 20 instruments. A combination of perceptual features shows the highest accuracy with around 90% and a reduced collection of selected features proves most efficient. The second experiment is similar to the first, but it processes musical passages to distinguish between instruments from four major instrument families. Spectral features are predominantly used for this task. Nineteen features combined from the Mel-frequency cepstral coefficients and the MPEG-7 standard achieve an accuracy of around 94%. In both experiments a multilayer perceptron shows the best generalisation for the test set.

The dataset from the second experiment is also used for a user interface that implements the most robust feature extraction technique and the fastest data analysis method. It detects the dominant instrument in a ten second long passage and annotates the calculated information in XML-format.