Reducing Interference with Phase Recovery in DNN-based Monaural Singing Voice Separation

Abstract

State-of-the-art methods for monaural singing voice separation consist in estimating the magnitude spectrum of the voice in the short-time Fourier transform (STFT) domain by means of deep neural networks (DNNs). The resulting magnitude estimate is then combined with the mixture's phase to retrieve the complex-valued STFT of the voice, which is further synthesized into a time-domain signal. However, when the sources overlap in time and frequency, the STFT phase of the voice differs from the mixture's phase, which results in interference and artifacts in the estimated signals. In this paper, we investigate on recent phase recovery algorithms that tackle this issue and can further enhance the separation quality. These algorithms exploit phase constraints that originate from a sinusoidal model or from consistency, a property that is a direct consequence of the STFT redundancy. Experiments conducted on real music songs show that those algorithms are efficient for reducing interference in the estimated voice compared to the baseline approach.

Phase‐aware subspace decomposition for single channel speech separation

“…As well, in this work, the importance of SS/enhancement of SD-based approaches when combined with the spectral phase has been pointed out. In fact, the phase information was proven to enhance speech quality when integrated into several SS processes (with NMF [15,16], with DNN [25,26,29], time-frequency masks [28]) but was never investigated in SD-based approaches. Hence, the proposed work is further support for phase importance for speech processing.…”

“…Hence, recent DNN-based approaches also exploit recent phase recovery algorithms that tackle this issue in order to further enhance the separation quality. To this end, phase-aware DNNs can exploit phase information in three main scenarios; the first one consists of two-stage approaches that are magnitude separation then phase enhancement [25].…”

Phase Recovery with Bregman Divergences for Audio Source Separation

“…Applying a nonnegative mask to the mixture's STFT results in assigning its phase to each isolated source. Even though this practice is common and yields satisfactory results, it is well established [8] that when sources overlap in the TF domain, using the mixture's phase induces residual interference and artifacts in the estimates. With the advent of deep learning, magnitudes can nowadays be estimated with a high accuracy, which outlines the need for more advanced phase recovery algorithms [9].…”

“…With the advent of deep learning, magnitudes can nowadays be estimated with a high accuracy, which outlines the need for more advanced phase recovery algorithms [9]. Consequently, a significant research effort has been put on phase recovery in DNN-based source separation, whether phase recovery algorithms are applied as a post-processing [8] or integrated within end-to-end systems for time-domain separation [10,11,12].…”

Online Spectrogram Inversion for Low-Latency Audio Source Separation

“…The work of P. Magron was conducted while he was with Tampere University and supported by the Academy of Finland, project no. 290190. focused on phase recovery in DNN-based source separation, whether phase recovery algorithms are applied as a postprocessing [12], [13] or integrated within end-to-end systems for time-domain separation [14], [15], [16], [17], [18].…”