David Gunawan

Iterative Phase Estimation for the Synthesis of Separated Sources From Single-Channel Mixtures

In this letter, we propose a novel method of refining the time-domain synthesis of individual source estimates from a single channel mixture. Employing a closed-loop architecture, the algorithm refines the synthesis of each source by iteratively estimating the phase of the sources, given the estimates of the source magnitude spectra and a single channel time-domain mixture. The performance of the algorithm is evaluated for harmonic musical mixtures, and considerable improvements to the synthesized estimates are obtained relative to phase binary masking, given accurate source magnitude spectra.

Sinusoidal Frequency Estimation based on the Time Derivative of the STFT Phase Response

This paper presents the phase derivative FFT (PDFFT)-a computationally efficient algorithm for estimating the frequency of a sinusoid from the short time Fourier transform (STFT). Upon obtaining initial coarse estimates from the FFT of a given frame, the PDFFT makes further refinement to the frequency estimate using only the time derivative of the phase response. The algorithm is derived and is shown to require only 4 multiplies per peak. Single frequencies in the presence of noise are resolved well, outperforming the commonly used quadratically interpolated FFT (QIFFT) method even with zero-padding. The algorithm is then used to separate two sinusoids of close frequency proximity that appear as a single peak in the magnitude spectrum

Spectro-temporal modeling of harmonic magnitude tracks for music source separation

Resolving overlapping harmonics remains a persistent and unsolved issue when separating individual sources from a single channel mixture of harmonic musical instruments. In this paper we present a novel method for resolving overlapping harmonics using a harmonic magnitude track prediction model, which exploits the spectro-temporal correlations that exist between instrument harmonics. The performance of the model is evaluated against other approaches [1], [2], [3], and is shown to provide better and more robust estimates of the harmonic magnitudes as a function of time.

Music source separation synthesis using Multiple Input Spectrogram Inversion

In this paper, we propose a novel method of refining the time-domain synthesis of individual source estimates from a single channel mixture. Employing a closed-loop architecture, the algorithm refines the synthesis of each source by iteratively estimating the phase of the sources, given the estimates of the source magnitude spectra and a single channel time-domain mixture. The performance of the algorithm is evaluated for harmonic musical mixtures, and considerable improvements to the synthesized estimates are obtained relative to phase binary masking, given accurate source magnitude spectra.

Musical Instrument Spectral Envelope Sensitivity

While a number of studies have explored discrimination thresholds for changes to the spectral envelope, the question of how sensitivity varies as a function of centre frequency and bandwidth to musical instruments has yet to be addressed. In this paper we conducted a 2AFC experiment to observe the discrimination thresholds of the trumpet, clarinet and viola for 14 different modifications of centre frequency and bandwidth. The results indicate that perceptual sensitivity has an SNR upper bound of 20 dB, governed by the first few harmonics and sensitivity does not really improve when extending the bandwidth any higher. However, sensitivity was found to decrease if changes were made only to the higher harmonics and continued to decrease as the distorted bandwidth was widened. The results are analyzed and discussed with respect to two other spectral envelope discrimination studies in the literature as well as what is predicted from a psychoacoustic model.