Speech Denoising Research Articles

Predictive denoising of speech in noise using deep neural networks

Deep neural networks have recently been used in several studies to improve speech intelligibility in noise. Here, we tested whether such networks could denoise speech in a predictive manner, which would be highly desirable for potential real-time applications. Training targets for the networks consisted in a mask (ideal binary mask or ideal ratio mask) for the last observed frame (non-predictive) and for one frame ahead in the future (predictive). Frame length was fixed at 48 ms. Training was performed on a target speaker with added speech-shaped noise, using about 25 min of training speech, at different signal to noise ratios ranging from −12 to 3 dB. A behavioral experiment was run to measure intelligibility of semantically unpredictable sentences in speech-shaped noise. For the behavioral experiment, target sentences were different from the learning sentences, and novel exemplars of speech-shaped noise were drawn on each trial. We observed intelligibility gains for both network architectures over a broad range of signal to noise ratios, with a maximum of 13.6 percentage points for the non-predictive network compared to 9.4 percentage points for the predictive network. This shows that a network may successfully be trained to denoise a specific speaker in a predictive manner.

Joint Denoising and Dereverberation Using Exemplar-Based Sparse Representations and Decaying Norm Constraint

Exemplar-based nonnegative models, where the noisy speech is decomposed as a sparse nonnegative linear combination of the speech and noise exemplars stored in a dictionary, have been successfully used for speech denoising. This paper extends this technique for the single-channel speech enhancement in noisy reverberant environments using a novel approximation of the noisy reverberant speech in the frequency domain and nonnegative matrix deconvolution. In the proposed model, the room impulse response RIR in the magnitude short-time Fourier transform domain is defined such that its decaying structure can also be estimated from the test data itself, whereas the existing models used a suboptimal binwise clamping procedure to impose such a decaying structure that does not hold in a typical RIR. This paper presents multiplicative updates for estimating the RIR, its decay, and the underlying anechoic speech and noise. The proposed model is evaluated on a synthetically created dataset created by convolving TIMIT recordings with RIRs measured from different rooms and varying speaker-and-microphone locations, and adding background noises taken from the CHiME corpus. Simulation results show that the proposed model results in a better RIR estimate over the existing model and improves various instrumental speech quality measures.

Speech Denoising Using Transform Domains in the Presence of Impulsive and Gaussian Noises

The speech denoising problem in the presence of mixed impulsive and Gaussian noises is investigated by exploiting transform domains. To that end, the proposed noise suppression scheme is a cascaded form consisting of an impulsive noise suppression module and a Gaussian noise suppression module. For the impulsive noise reduction subsystem, in this paper, the noise is sparsely represented by the time domain, whereas short-time Fourier transform, wavelet transform, and wavelet synchrosqueezed transform are studied to provide sparse representations for the speech. By utilizing the transform domains, the speech recovery and the impulsive noise suppression are simultaneously achieved under an optimization framework. Subsequently, the alternating direction method of multipliers is used to solve $\ell _{1}$ -norm constrained optimization. In the Gaussian noise reduction subsystem, the Gaussian noise is suppressed by the famous Wiener filter in the transform domains as well. Numerical studies, including simulations and real data analysis, demonstrate the superior performance of the proposed scheme.

Compression and denoising of speech transmission using Daubechies wavelet family

Over the past few years there has been much work done on compression and enhancement of speech signals. Efficient speech compression techniques are becoming hectic with the fast growth of data and multimedia services. In this paper, wavelet-based analysis, processing and compression of multimedia signals like speech are described and implemented. A wavelet-based speech compression and speech denoising techniques obtained by using hard and soft thresholding algorithm are presented. Daubechies wavelets have been implemented on a test recorded speech signal for compression and denoising purposes. It has been found that the proposed work shows better results as compared to the already existing technique. Moreover, soft thresholding is preferable over hard thresholding with higher PSNR and less MSE. The results obtained have been found satisfactory as compared to existing state-of-the-art techniques.

Compression and denoising of speech transmission using Daubechies wavelet family

Over the past few years there has been much work done on compression and enhancement of speech signals. Efficient speech compression techniques are becoming hectic with the fast growth of data and multimedia services. In this paper, wavelet-based analysis, processing and compression of multimedia signals like speech are described and implemented. A wavelet-based speech compression and speech denoising techniques obtained by using hard and soft thresholding algorithm are presented. Daubechies wavelets have been implemented on a test recorded speech signal for compression and denoising purposes. It has been found that the proposed work shows better results as compared to the already existing technique. Moreover, soft thresholding is preferable over hard thresholding with higher PSNR and less MSE. The results obtained have been found satisfactory as compared to existing state-of-the-art techniques.

Estimation of Noise Magnitude for Speech Denoising Using Minima-Controlled-Recursive-Averaging Algorithm Adapted by Harmonic Properties

The accuracy of noise estimation is important for the performance of a speech denoising system. Most noise estimators suffer from either overestimation or underestimation on the noise level. An overestimate on noise magnitude will cause serious speech distortion for speech denoising. Conversely, a great quantity of residual noise will occur when the noise magnitude is underestimated. Accurately estimating noise magnitude is important for speech denoising. This study proposes employing variable segment length for noise tracking and variable thresholds for the determination of speech presence probability, resulting in the performance improvement for a minima-controlled-recursive-averaging (MCRA) algorithm in noise estimation. Initially, the fundamental frequency was estimated to determine whether a frame is a vowel. In the case of a vowel frame, the increment of segment lengths and the decrement of threshold for speech presence were performed which resulted in underestimating the level of noise magnitude. Accordingly, the speech distortion is reduced in denoised speech. On the contrary, the segment length decreases rapidly in noise-dominant regions. This enables the noise estimate to update quickly and the noise variation to track well, yielding interference noise being removed effectively through the process of speech denoising. Experimental results show that the proposed approach has been effective in improving the performance of the MCRA algorithm by preserving the weak vowels and consonants. The denoising performance is therefore improved.

Greedy double sparse dictionary learning for sparse representation of speech signals

This paper proposes a greedy double sparse (DS) dictionary learning algorithm for speech signals, where the dictionary is the product of a predefined base dictionary, and a sparse matrix. Exploiting the DS structure, we show that the dictionary can be learned efficiently in the coefficient domain rather than the signal domain. It is achieved by modifying the objective function such that all the matrices involved in the coefficient domain are either sparse or near-sparse, thus making the dictionary update stage fast. The dictionary is learned on frames extracted from a speech signal using a hierarchical subset selection approach. Here, each dictionary atom is a training speech frame, chosen in accordance to its energy contribution for representing all other training speech frames. In other words, dictionary atoms are encouraged to be close to the training signals that uses them in their decomposition. After each atom update the modified residual serves as the new training data, thus the information learned by the previous atoms guides the update of subsequent dictionary atoms. In addition, we have shown that for a suitable choice of the base dictionary, storage efficiency of the DS dictionary can be further improved. Finally, the efficiency of the proposed method is demonstrated for the problem of speech representation and speech denoising.

Simultaneous speech coding and de-noising in a dictionary based quantized CS framework

Speech compression or speech coding is inevitable for effective communication of speech signals in resource limited scenarios and researcher’s have been working on achieving lower and lower transmission bit rates (BR) without much compromise on the quality of speech. Medium BR hybrid speech coding schemes have gained much interest in the recent years with most of them based on CELP, the basic medium bit-rate coding scheme. In this work, we provide an insight to the capabilities of compressive sensing (CS) in speech processing and propose a novel idea in the quantized framework. Three major aspects demonstrated in this paper are (1) Inherent de-noising of noisy speech by the CS based coder along with compression (2) Quantization of CS measurements to achieve medium transmission bit-rates and (3) Enhancement of quality and compression performance of the coder with better sparse representations of speech using dictionaries. The results indicate that the proposed scheme offers better compression in comparison with basic Gaussian codebook CELP. The CS scheme has the added advantage of inherent noise suppression and provides more robustness to background noise in comparison with parameter extraction based medium bit-rate speech coding systems.

Compositional model for speech denoising based on source/filter speech representation and smoothness/sparseness noise constraints

We present a speech denoising algorithm based on a regularized non-negative matrix factorization (NMF), in which several constraints are defined to describe the background noise in a generic way. The observed spectrogram is decomposed into four signal contributions: the voiced speech source and three generic types of noise. The speech signal is represented by a source/filter model which captures only voiced speech, and where the filter bases are trained on a database of individual phonemes, resulting in a small dictionary of phoneme envelopes. The three remaining terms represent the background noise as a sum of three different types of noise (smooth noise, impulsive noise and pitched noise), where each type of noise is characterized individually by imposing specific spectro-temporal constraints, based on sparseness and smoothness restrictions. The method was evaluated on the 3rd CHiME Speech Separation and Recognition Challenge development dataset and compared with conventional semi-supervised NMF with sparse activations. Our experiments show that, with a similar number of bases, source/filter modeling of speech in conjunction with the proposed noise constraints produces better separation results than sparse training of speech bases, even though the system is only designed for voiced speech and the results may still not be practical for many applications.

Joint Optimization of Masks and Deep Recurrent Neural Networks for Monaural Source Separation

Monaural source separation is important for many real world applications. It is challenging because, with only a single channel of information available, without any constraints, an infinite number of solutions are possible. In this paper, we explore joint optimization of masking functions and deep recurrent neural networks for monaural source separation tasks, including monaural speech separation, monaural singing voice separation, and speech denoising. The joint optimization of the deep recurrent neural networks with an extra masking layer enforces a reconstruction constraint. Moreover, we explore a discriminative criterion for training neural networks to further enhance the separation performance. We evaluate the proposed system on the TSP, MIR-1K, and TIMIT datasets for speech separation, singing voice separation, and speech denoising tasks, respectively. Our approaches achieve 2.30--4.98 dB SDR gain compared to NMF models in the speech separation task, 2.30--2.48 dB GNSDR gain and 4.32--5.42 dB GSIR gain compared to existing models in the singing voice separation task, and outperform NMF and DNN baselines in the speech denoising task.

A Hybrid Nonlinear ANC for Speech Recovery Using Both Bone- and Air-Conducted Measurements

<div class=""abs_img""> <img src=""[disp_template_path]/JRM/abst-image/00270005/08.jpg"" width=""300"" /> Denoising nonlinear filter</div> Speech recovery in the presence of very harsh noise is calling for R&D that take approaches different from those established to date, as the conventional systems and algorithms for speech denoising may suffer from serious performance degradation. In this paper, we propose a hybrid nonlinear adaptive noise canceller (ANC) to perform the speech enhancement task using both bone- and air-conducted measurements. In the proposed ANC, the bone-conducted speech serves as a reference signal while the air-conducted measurement with very large additive noise is adopted as a primary noise. A Volterra filter and a functional link artificial neural network (FLANN) are placed in parallel, forming a hybrid nonlinear ANC. Simulations using real bone- and air-conducted speech measurements are provided to demonstrate that the proposed system outperforms ANCs equipped with FIR filter or Volterra filter or FLANN alone. </span>

Optimization of learned dictionary for sparse coding in speech processing

As a promising technique, sparse coding has been widely used for the analysis, representation, compression, denoising and separation of speech. This technique needs a good dictionary which contains atoms to represent speech signals. Although many methods have been proposed to learn such a dictionary, there are still two problems. First, unimportant atoms bring a heavy computational load to sparse decomposition and reconstruction, which prevents sparse coding from real-time application. Second, in speech denoising and separation, harmful atoms have no or ignorable contributions to reducing the sparsity degree but increase the source confusion, resulting in severe distortions. To solve these two problems, we first analyze the inherent assumptions of sparse coding and show that distortion can be caused if the assumptions do not hold true. Next, we propose two methods to optimize a given dictionary by removing unimportant atoms and harmful atoms, respectively. Experiments show that the proposed methods can further improve the performance of dictionaries.

Spectrum enhancement with sparse coding for robust speech recognition

Recently, a trend in speech recognition is to introduce sparse coding for noise robustness. Although several methods have been proposed, the performance of sparse coding in speech denoising is not so optimistic. One assumption with sparse coding is that the representation of speech over the speech dictionary is sparse, while that of the noise is dense. This assumption is obviously not sustained in the speech denoising scenario. Many noises are also sparse over the speech dictionary. In such a condition, the representation of noisy speech still contains noise components, resulting in degraded performance. To solve this problem, we first analyze the assumption of sparse coding and then propose a novel method to enhance speech spectrum. This method first finds out the atoms which represent the noise sparsely, and then selectively ignores them in the reconstruction of speech to reduce the residual noise. Speech features are then extracted from the enhanced spectrum for speech recognition. Experimental results show that the proposed method can improve the noise robustness of a speech recognition system substantially.

Dictionary evaluation and optimization for sparse coding based speech processing

Recently, sparse coding has attracted considerable attention in speech processing. As a promising technique, sparse coding can be widely used for analysis, representation, compression, denoising and separation of speech. To represent signals accurately and sparsely, a good dictionary which contains elemental signals is preferred and many methods have been proposed to learn such a dictionary. However, there is a lack of reasonable evaluation methods to judge whether a dictionary is good enough. To solve this problem, we define a group of measures for dictionary evaluation. These measures not only address sparseness and reconstruction error of signal representation, but also consider denoising and separating performance. We show how to evaluate dictionaries with these measures, and further propose two methods to optimize dictionaries by improving relative measures. The first method improves the efficiency of sparse coding by removing unimportant atoms; the second one improves denoising performance of dictionaries by removing harmful atoms. Experimental results show that the measures can provide reasonable evaluations and the proposed methods for optimization can further improve given dictionaries.

BionicWavelet Based Denoising Using Source Separation

We consider the problem of speech denoising using source separation. In this study we have proposed a hybrid technique that consists in applying in the first step, the Bionic Wavelet Transform (BWT) to two different mixtures of the same speech signal with noise. This speech signal is corrupted by a Gaussian white noise with two different values of the Signal to Noise Ratio (SNR) in order to obtain those two mixtures. The second step consists in computing the entropy of each bionic wavelet coefficient and finds the two subbands having the minimal entropy. Those two subbands are used to estimate the separation matrix of the speech signal from noise by using the source separation. Our proposed technique is evaluated by comparing it to the denoising technique based on source separation in time domain.

English

This paper investigates the improvement of Telugu speech quality in terms of six objective quality measures using Discrete Wavelet Transform and proposes two Hybrid thresholding methods which are formed by combining soft and Improved thresholding methods with Modified Improved thresholding method. The performance of the new Hybrid methods is compared with the other thresholding methods. It is observed that the new proposed scheme yields better results when applied to Telugu noisy speech signals with low SNR (0dB) conditions. In this method, noisy speech signal is divided in to overlapping frames and each frame is windowed using hamming window. The windowed speech blocks are applied to the wavelet based speech enhancement algorithm and the enhanced speech is reconstructed in its time domain. For denoising the Telugu speech signal, various techniques like hard, soft, improved, modified improved and hybrid thresholding methods are used. Analysis is done using daubechies and symlets wavelets with different white Gaussian noise environments. Six Objective quality measures are considered in this study to test the performance of the algorithm for enhanced Telugu speech quality and compared. Hybrid thresholding methods perform better than hard, soft, improved and modified improved thresholding methods for wavelet based speech denoising.

An Improvement Based on Wavelet Transform Speech Denoising Threshold Function

The wavelet threshold denoising method is widely researched and used, the hard threshold and soft threshold which exist different advantages and disadvantages are two kinds commonly used threshold function; the threshold function was improved between hard threshold and soft threshold by using the exponential function. The smooth curve is monotonically increasing, no discontinuity and thus will not produce oscillation. Overcoming shortcomings that the hard threshold function and soft threshold function to be zero in wavelet coefficient is small and soft threshold exist constant deviation, and when the absolute value is greater than the threshold coefficient, soft threshold function is disposed with the same contraction method. After the simulation results prove that compared with traditional speech enhancement algorithms, the new threshold function has been improved to a certain extent.

Empirical Mode Decomposition and Normal Shrink Tresholding for Speech Denoising

In this paper a signal denoising scheme based on Empirical mode decomposition (EMD) is presented. the noisy signal is decomposed in an adaptive manner by the EMD algorithm which allows to obtain intrinsic oscillatory called Intrinsic mode functions (IMFs)component by a process called sifting process. The basic principle of the method is to decompose a speech signal corrupted by additive white Gaussian random noise into segments each frame is categorised as either signal-dominant or noise-dominant then reconstruct the signal with IMFs signal dominant frame previously filtered or thresholded.It is shown, on the basis of intensivesimulations that EMD improves the signal to noise ratio and address the problem of signal degradation. The denoising method is applied to real signal with different noise levels and the results compared to Winner and universal threshold of DONOHO and JOHNSTONE [11] with soft and hard tresholding.Theeffect of level noise value on the performances of the proposed denoising is analysed.

Speech Denoising via Low-Rank and Sparse Matrix Decomposition

Speech Denoising via Low-Rank and Sparse Matrix Decomposition

Speech enhancement using empirical mode decomposition and the Teager–Kaiser energy operator

In this paper a speech denoising strategy based on time adaptive thresholding of intrinsic modes functions (IMFs) of the signal, extracted by empirical mode decomposition (EMD), is introduced. The denoised signal is reconstructed by the superposition of its adaptive thresholded IMFs. Adaptive thresholds are estimated using the Teager-Kaiser energy operator (TKEO) of signal IMFs. More precisely, TKEO identifies the type of frame by expanding differences between speech and non-speech frames in each IMF. Based on the EMD, the proposed speech denoising scheme is a fully data-driven approach. The method is tested on speech signals with different noise levels and the results are compared to EMD-shrinkage and wavelet transform (WT) coupled with TKEO. Speech enhancement performance is evaluated using output signal to noise ratio (SNR) and perceptual evaluation of speech quality (PESQ) measure. Based on the analyzed speech signals, the proposed enhancement scheme performs better than WT-TKEO and EMD-shrinkage approaches in terms of output SNR and PESQ. The noise is greatly reduced using time-adaptive thresholding than universal thresholding. The study is limited to signals corrupted by additive white Gaussian noise.