Independent deeply learned matrix analysis with automatic selection of stable microphone-wise update and fast sourcewise update of demixing matrix

Makishima, Naoki; Mitsui, Yoshiki; Takamune, Norihiro; Kitamura, Daichi; Saruwatari, Hiroshi; Takahashi, Yu; Kondo, Kazunobu

doi:10.1016/j.sigpro.2020.107753

Cited by 5 publications

(6 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We explain that the problem of minimizing L with respect to W i when R jn is kept fixed can be solved by an algorithm called vectorwise coordinate descent (VCD) [15], which has been developed for the conventional IPSDTA [17]. VCD is a block coordinate descent method that cyclically update w in for each i = 1, .…”

Section: A Optimization Of Demixing Matrix W Imentioning

confidence: 99%

See 1 more Smart Citation

Independent Deeply Learned Tensor Analysis for Determined Audio Source Separation

Narisawa

Ikeshita

Takamune

et al. 2021

2021 29th European Signal Processing Conference (EUSIPCO)

Self Cite

View full text Add to dashboard Cite

We address the determined audio source separation problem in the time-frequency domain. In independent deeply learned matrix analysis (IDLMA), it is assumed that the interfrequency correlation of each source spectrum is zero, which is inappropriate for modeling nonstationary signals such as music signals. To account for the correlation between frequencies, independent positive semidefinite tensor analysis has been proposed. This unsupervised (blind) method, however, severely restrict the structure of frequency covariance matrices (FCMs) to reduce the number of model parameters. As an extension of these conventional approaches, we here propose a supervised method that models FCMs using deep neural networks (DNNs). It is difficult to directly infer FCMs using DNNs. Therefore, we also propose a new FCM model represented as a convex combination of a diagonal FCM and a rank-1 FCM. Our FCM model is flexible enough to not only consider inter-frequency correlation, but also capture the dynamics of time-varying FCMs of nonstationary signals. We infer the proposed FCMs using two DNNs: DNN for power spectrum estimation and DNN for time-domain signal estimation. An experimental result of separating music signals shows that the proposed method provides higher separation performance than IDLMA.

show abstract

Section: A Optimization Of Demixing Matrix W Imentioning

confidence: 99%

“…As a supervised extension of ILRMA as well as FDICA, independent deeply learned matrix analysis (IDLMA) [14], [15] has been proposed. Instead of using NMF, in IDLMA, a pretrained deep neural network (DNN) is used to model a source power spectrum.…”

Section: Introductionmentioning

confidence: 99%

Independent Deeply Learned Tensor Analysis for Determined Audio Source Separation

Narisawa

Ikeshita

Takamune

et al. 2021

2021 29th European Signal Processing Conference (EUSIPCO)

Self Cite

View full text Add to dashboard Cite

show abstract

“…More recently, deep neural network (DNN) is utilized to model the source spectral characteristics [13,14,15,16,17,18] given its powerful modeling ability. In [13,14,15,16], the supervised learning of the source spectrogram is presented, which is further combined with the ILRMA-based blind estimation of the demixing matrix and is referred as independent deeply learned matrix analysis (IDLMA). Due to the complexity of directly estimating the source spectrogram, the DNN-based supervised update of the source variance is then incorporated into the framework of IVA [17].…”

Section: Introductionmentioning

confidence: 99%

“…Taking advantage of the iterative source steering updates, the work in [18] proposes to determine the source model parameters by backing propagate the permutation invariant separation loss through multiple iterations of IVA. These DNN models [13,14,15,16,17,18] can accurately estimate the parameters of the given source model. However, they need to update the source model parameters with DNN iteratively, leading to the expensive computational cost and the mismatch between training and testing data.…”

Section: Introductionmentioning

confidence: 99%

A Novel Blind Source Separation Framework Towards Maximum Signal-To-Interference Ratio

Gu¹,

Cheng²,

Yao³

et al. 2021

Preprint

View full text Add to dashboard Cite

In this paper, the optimal source model for the independent vector analysis (IVA) algorithm towards maximizing the output signal-to-interference ratio (SIR) is mathematically derived, and the corresponding optimal weighted covariance matrix is proved to be the covariance matrix of interference signals. A new algorithm framework called minimum variance IVA (MVIVA) is further proposed, where the deep neural network-based estimation of the interference covariance matrix is combined with the IVA-based estimation of the demixing matrix. Experimental results show the superiority of the proposed source model, and the MVIVA algorithm outperforms the original IVA algorithm by 9.6 dB in SIR and 5.8 dB in signal-to-distortion ration (SDR) on average.

show abstract

“…We aimed to reduce bleeding sound in a fully blind manner, namely, the spatial locations of sources and microphones are unknown. We also did not use supervision of sources, such as solo-played music datasets, to avoid the mismatch between training and test data; thus, supervised deep-neural-networkbased approaches [19], [20], [21], [22], [23] are out of the scope of this paper. We propose a phase-insensitive method for blind bleeding-sound reduction, which is a modification of TCNMF: we introduce an a priori generative model for diagonal and off-diagonal elements of the frequency-wise mixing matrix to model relative leakage levels of bleeding sounds.…”

Section: Introductionmentioning

confidence: 99%

Prior Distribution Design for Music Bleeding-Sound Reduction Based on Nonnegative Matrix Factorization

Mizobuchi,

Kitamura,

Nakamura

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

When we place microphones close to a sound source near other sources in audio recording, the obtained audio signal includes undesired sound from the other sources, which is often called cross-talk or bleeding sound. For many audio applications including onstage sound reinforcement and sound editing after a live performance, it is important to reduce the bleeding sound in each recorded signal. However, since microphones are spatially apart from each other in this situation, typical phase-aware blind source separation (BSS) methods cannot be used. We propose a phase-insensitive method for blind bleeding-sound reduction. This method is based on time-channel nonnegative matrix factorization, which is a BSS method using only amplitude spectrograms. With the proposed method, we introduce the gamma-distribution-based prior for leakage levels of bleeding sounds. Its optimization can be interpreted as maximum a posteriori estimation. The experimental results of music bleedingsound reduction indicate that the proposed method is more effective for bleeding-sound reduction of music signals compared with other BSS methods.

show abstract

Independent deeply learned matrix analysis with automatic selection of stable microphone-wise update and fast sourcewise update of demixing matrix

Cited by 5 publications

References 29 publications

Independent Deeply Learned Tensor Analysis for Determined Audio Source Separation

Independent Deeply Learned Tensor Analysis for Determined Audio Source Separation

A Novel Blind Source Separation Framework Towards Maximum Signal-To-Interference Ratio

Prior Distribution Design for Music Bleeding-Sound Reduction Based on Nonnegative Matrix Factorization

Contact Info

Product

Resources

About