A novel approach to solve the single-channel source separation (SCSS) problem is presented. Most existing supervised SCSS methods resort exclusively to the independence waveform criteria as exemplified by training the prior information before the separation process. This poses a significant limiting factor to the applicability of these methods to real problem. Our proposed method does not require training knowledge for separating the mixture and it is based on decomposing the mixture into a series of oscillatory components termed as the intrinsic mode functions (IMFs). We show, in this paper, that the IMFs have several desirable properties unique to SCSS problem and how these properties can be advantaged to relax the constraints posed by the problem. In addition, we have derived a novel sparse non-negative matrix factorization to estimate the spectral bases and temporal codes of the sources. The proposed algorithm is a more complete and efficient approach to matrix factorization where a generalized criterion for variable sparseness is imposed onto the solution. Experimental testing has been conducted to show that the proposed method gives superior performance over other existing approaches.

Single-Channel Source Separation Using EMD-Subband Variable Regularized Sparse Features

Cepstral features and their derivatives have been successfully used in speaker recognitions tasks over the past two decades Recently new features are emerging that make use of information that is not contained in cepstral features or their derivative, namely the phase spectra and instantaneous frequency There are numerous approaches to capture instantaneous amplitude and frequency variations in the signal, resulting in an amplitude and frequency modulation (AM-FM) feature set While the addition of modulation features has been shown to improve speech recognition systems based on cepstral features, their use in speaker recognition is new and has not been fully explored This paper outlines some of these emerging features and their extraction methods for use in speaker recognition tasks The paper also talks about the recently proposed Empirical Mode Decomposition (EMD) as an alternative method of extracting instantaneous frequency

Emerging features for speaker recognition

This paper introduces a new voiced speech denoising approach based on the empirical mode decomposition (EMD) associated to an appropriate sifting process. Noisy signal is decomposed adaptively into intrinsic oscillatory components called (IMFs). Since, the energy of a voiced speech signal is distributed over low and medium frequencies, the obtained lower order IMFs (high frequency components) are expected to be highly contaminated by noise. However, the last IMFs (low and medium frequency components) corresponding to the most structures of the signal, contain very low noise levels. Therefore the filtering of the low order IMFs may introduce a signal distortion rather than reducing noise. The principle of the proposed method is then based on filtering only the first IMFs considered very noisy by the adaptive center weighted average (ACWA) filter. A criterion based on the IMFss energies is used to select the very noisy IMFs that must be filtered. The denoising proposed method is applied successfully to voiced speech signal corrupted with additive white Gaussian noise. The reported results obtained for different noise levels, demonstrate the efficiency of the proposed method for reducing noise and its superiority over other denoising methods considered for comparison.

A new EMD denoising approach dedicated to voiced speech signals

This study presents a speech filtering method exploiting the combined effects of the empirical mode decomposition (EMD) and the local statistics of the speech signal using the adaptive centre weighted average (ACWA) filter. The novelty lies in incorporating the frame class (voiced/unvoiced) in the conventional filtering using the EMD and the ACWA filter. The speech signal is segmented into frames and each one is broken down by the EMD into a finite number of intrinsic mode functions (IMFs). The number of filtered IMFs depends on whether the frame is voiced or unvoiced. An energy criterion is used to identify voiced frames while a stationarity index distinguishes between unvoiced and transient sequences. Reported results obtained on signals corrupted by additive noise (white, F16, factory) show that the proposed filtering in line with the frame class is very effective in removing noise components from noisy speech signal. Compared with filtering results of the wavelet, the ACWA, and the EMD-ACWA methods, the proposed technique gives much better results in terms of average segmental signal-to-noise ratio and listening quality based on perceptual evaluation speech quality score.

https://ietresearch.onlinelibrary.wiley.com/doi/epdf/10.1049/iet-spr.2013.0425

Voiced/unvoiced speech classification-based adaptive filtering of decomposed empirical modes for speech enhancement

In this paper a new method for voiced speech enhancement combining the Empirical Mode Decomposition (EMD) and the Adaptive Center Weighted Average (ACWA) filter is introduced. Noisy signal is decomposed adaptively into intrinsic oscillatory components called Intrinsic Mode Functions (IMFs). Since voiced speech structure is mostly distributed on both medium and low frequencies, the shorter scale IMFs of the noisy signal are beneath noise, however the longer scale ones are less noisy. Therefore, the main idea of the proposed approach is to only filter the shorter scale IMFs, and to keep the longer scale ones unchanged. In fact, the filtering of longer scale IMFs will introduce distortion rather than reducing noise. The denoising method is applied to several voiced speech signals with different noise levels and the results are compared with wavelet approach, ACWA filter and EMD–ACWA (filtering of all IMFs using ACWA filter). Relying on exhaustive simulations, we show the efficiency of the proposed method for reducing noise and its superiority over other denoising methods, i.e. to improve Signal-to-Noise Ratio (SNR), and to offer better listening quality based on a Perceptual Evaluation of Speech Quality (PESQ). The present study is limited to signals corrupted by additive white Gaussian noise.

/pdf/voiced-speech-enhancement-based-on-adaptive-filtering-of-1wyv5yp68i.pdf

Voiced speech enhancement based on adaptive filtering of selected intrinsic mode functions

This paper introduces a new speech enhancement method using soft thresholding with a Discrete Cosine Transform (DCT) and Empirical Mode Decomposition (EMD) based hybrid algorithm. Soft thresholding for DCT-enhancement is a powerful method for enhancing the noisy speech signal in a wide range of signal-to-noise ratios (SNRs). However, due to the thresholding criteria a significant amount of noise is left in the enhanced signal. EMD is applied here to remove the remaining noise components. Due to the frequency characteristics of the intrinsic mode functions (IMFs), the noise components are mainly centered in the lower order IMFs. Therefore, it is possible to successfully identify and remove the remaining noise. The experimental results show that the proposed hybrid method is significantly more effective in removing the noise components from the noisy speech signal; thus giving better results in output SNR and quality compared to recently reported techniques.

Speech enhancement using soft thresholding with DCT-EMD based hybrid algorithm

This paper introduces a novel speech enhancement method based on Empirical Mode Decomposition (EMD) and softthresholding algorithms. A modified soft thresholding strategy is adapted to the intrinsic mode functions (IMF) of the noisy speech. Due to the characteristics of EMD, each obtained IMF of the noisy signal will have different noise and speech energy distribution, thus will have a different noise variance. Based on this specific noise variance, by applying the proposed thresholding algorithm to each IMF separately, it is possible to effectively extract the existing noise components. The experimental results suggest that the proposed method is significantly more effective in removing the noise components from the noisy speech signal compared to recently reported techniques. The significantly better SNR improvement and the speech quality prove the superiority of the proposed algorithm. Index Terms: speech enhancement, empirical mode decomposition, soft-thresholding

EMD based Soft-Thresholding for Speech Enhancement

This paper presents a two-stage soft thresholding algorithm based on discrete cosine transform (DCT) and empirical mode decomposition (EMD). In the first stage, noisy speech is decomposed into eight frequency bands and a specific noise variance is calculated for each one. Based on this variance, each band is denoised using soft thresholding in DCT domain. The remaining noise is eliminated in the second stage through a time domain soft thresholding strategy adapted to the intrinsic mode functions (IMFs) derived by applying EMD on the signal obtained from the first stage processing. Significantly better SNR improvement and perceptual speech quality results for different noise types prove the superiority of the proposed algorithm over recently reported techniques.

/pdf/subband-dct-and-emd-based-hybrid-soft-thresholding-for-2cmng0mi9g.pdf

Erhan Deger

Papers

Speech enhancement using soft thresholding with DCT-EMD based hybrid algorithm

EMD based Soft-Thresholding for Speech Enhancement

Subband DCT and EMD Based Hybrid Soft Thresholding for Speech Enhancement