Most research in musical instrument identification has focused on labeling isolated samples or solo phrases. A robust instrument identification system capable of dealing with polytimbral recordings of instruments remains a necessity in music information retrieval. Experiments are described which evaluate the ground truth of ADRess as a sound source separation technique used as a preprocess to automatic musical instrument identification. The ground truth experiments are based on a number of basic acoustic features, while using a Gaussian Mixture Model as the classification algorithm. Using all 44 acoustic feature dimensions, successful identification rates are achieved.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.