Extraction of weak background transients from audio signals

Extracting a weak audio signal buried below a stronger one is a difficult task that may be encountered in forensic applications. Extraction of the background signal is made difficult by the signal-to-noise ratio that is clearly negative. Apart from this difficulty, the problem would look like the problem of separating different components from an audio signal. We investigate the possibility of using separation techniques for this extraction, based upon harmonic models of the stronger signal. We compare theoretically and practically two approaches: in the first one, estimation of harmonic models followed by subtraction of the harmonic signal, in the second one, high resolution tracking of damped sinusoids and synthesis of each component after sorting the individual patterns.