Hello, and welcome to Paper of the Day (Po’D): Transients Detection Edition. Today’s paper comes from the March 2010 issue of IEEE Transactions on Audio, Speech, and Language Processing, a special issue devoted to signal models and representations of environmental sounds: V. Bruni, S. Marconi, and D. Vitulano, “Time-scale Atoms Chains for Transients Detection in Audio Signals,” IEEE TASLP, vol. 18, no. 3, pp. 420-433, Mar. 2010.
In this paper, the authors propose a means for locating transients in audio signals. They first propose a model of an audio signal as a superposition of “three distinct components:” a tonal, or locally stationary, component; a transient component; and a stochastic component that is unexplained by the tonal and transient components. The authors suggest modeling an individual transient as a piecewise linear function having at least three singular points, corresponding to onset, end of attack, and end of decay. Locating these singular points then is equivalent to detecting and delimiting transients. The authors propose to do so by following the maximum moduli of continuous-time wavelet transforms
from atom “chains” going from the smallest scale to large scales. For a portion of a signal that is tonal, the maximum moduli will decay faster across scales; but for a portion with a transient, it will decay much slower due to the presence of singularities. The authors claim that their algorithm does not require any interaction from the user in order to tune it.
Unfortunately, I cannot find any web page with sound examples, or testable software, accompanying this research, so I am not sure about the usefulness of this work. I did find this web page, but it does not help much. Reproducible research must be reproducible.