Marek Dziubinski

Estimation of musical sound separation algorithm effectiveness employing neural networks

Blind separation of musical sounds contained in sound mixtures is a challenging and difficult task. It is due to the fact that in Western music, mixed harmonic sources may be correlated with each other, i.e. their harmonic partials might be overlapping in the frequency domain if the signals remain in harmonic relation. Evaluation of the separation results is also problematic, since analysis of the energy-based error between the original signals used for mixing and the separated ones, in some cases, do not correspond with perceptual evaluation results. In this paper, four separation algorithms, engineered by the Authors, are presented. Then, musical instrument sound identification based on artificial neural networks is performed as a means of evaluating the performance of the separation algorithms. Results are discussed and conclusions are derived.

Octave Error Immune and Instantaneous Pitch Detection Algorithm

Abstract The aim of this article is to present an octave error optimized pitch detection algorithm based on spectral analysis. The proposed algorithm is effective for signals with strong harmonic content, as well as for nearly sinusoidal ones. In addition, as an extension to the presented octave error optimized algorithm, a method of estimating instantaneous pitch is described. Experiments and estimation accuracy tests in terms of octave errors were performed on a variety of musical instruments (i.e., 567 sounds played on acoustic instruments with various articulations and dynamics, with fundamental frequencies ranging from 34 Hz up to 1700 Hz, were processed). Fine pitch error tests of the instantaneous pitch estimation algorithm were performed for 4,000 different synthetic signals, with frequencies ranging from 50 Hz to 4000 Hz, including both clean signals and signals contaminated with noise. Results exemplifying the main issues of both engineered algorithms are shown. In addition, a performance comparison between the engineered algorithm and algorithms from the Wavesurfer software is presented.

Intelligent algorithms for optical track audio restoration

The Unpredictability Measure computation algorithm applied to psychoacoustic model-based broadband noise attenuation is discussed. A learning decision algorithm based on a neural network is employed for determining audio signal useful components acting as maskers of the spectral components classified as noise. An iterative algorithm for calculating the sound masking pattern is presented. The routines for precise extraction of sinusoidal components from sound spectrum were examined, such as estimation of pitch variations in the optical track audio affected by parasitic frequency modulation. The results obtained employing proposed intelligent signal processing algorithms will be presented and discussed in the paper.

Noise Reduction in Audio Employing Spectral Unpredictability Measure and Neural Net

Improvements of the recently presented noise reduction algorithm, based on perceptual coding of audio are revealed. Enhancements of the spectral Unpredictability Measure parameter calculation, which is one of the significant elements in the applied psychoacoustic model are discussed. A learning decision algorithm based on a neural network is employed for determining input signal useful components acting as maskers of the spectral components classified as noise. A new iterative algorithm for calculating the masking pattern is presented. The results of experiments carried out employing the modified algorithm are discussed and conclusions are added.

Evaluation of the Separation Algorithm Performance Employing ANNs

The objective of the presented study is to show that it is possible to effectively separate harmonic sounds from musical sound mixtures for the purpose of automatic sounds recognition, without any prior knowledge of the mixed instruments. It has also been shown that a properly trained ANN enables to reliably validate separation results of mixed musical instrument sounds, and the validation corresponds with subjective perception of the separated sounds quality. A comparison between the results obtained with the use of the ANN-based recognition, subjective evaluation of the separation performance and the energy-based evaluation is provided.

Intelligent algorithms for movie sound tracks restoration

Two algorithms for movie sound tracks restoration are discussed in the paper. The first algorithm is the unpredictability measure computation applied to the psychoacoustic model-based broadband noise attenuation. A learning decision algorithm, based on a neural network, is employed for determining useful audio signal components acting as maskers of the noisy spectral parts. An application of the rough set decision system to this task is also considered. An iterative method for calculating the sound masking pattern is presented. The second of presented algorithms is the routine for precise evaluation of parasite frequency modulations (wow) utilizing sinusoidal components extracted from the sound spectrum. The results obtained employing proposed intelligent signal processing algorithms, as well as the relationship between both routines, will be presented and discussed in the paper.