1 Introduction

In everyday life, target speech reaching the human auditory system is always accompanied by many other sound sources such as musical sounds, voices of other people speaking and shouting, environmental sounds and man-made sounds. In the last few decades, many researchers developed speech segregation system to segregate the target speech effectively from the other sound sources, unfortunately, there is no such system exist till now. The speech segregation systems find a lot of potential applications including noise robust automatic speech and speaker recognition, music transcription, digital content management, and mobile communication. There are applications which use only one microphone to record the target speech, for example, voice communication, audio retrieval and hearing aids. Segregating the target speech from one microphone recording of the speech is named as monaural speech segregation and it is one of the most challenging problems in speech signal processing. Although monaural speech segregation is a challenging problem, humans have the inherent capability to focus and understand the target speech recorded using a single microphone in spite of disturbances caused by other sound sources. In the last few decades, research work has been carried on more than hundreds of articles on monaural approaches, and some of the techniques in monaural speech segregation are spectral subtraction (Boll 1979; Hu and Wang 2011; Therese and Lingam 2017), subspace analysis (Ephraim and Trees 1995), model-based approaches (Ellis and Weiss 2006), and feature-based approach (Hu and Wang 2006; Phapatanaburi et al. 2017). Spectral subtraction is the conventional speech enhancement method for reducing the additive noise. This method reduces the noise by subtracting the estimated target speech spectrum from the noisy speech spectrum in the frequency domain (Boll 1979). The main drawback of this method is that it contains some spectral artifacts called remnant noise which irritates the human ear (Boll 1979; Hu and Wang 2011). In subspace analysis, the vector space of noisy speech will be decomposed into speech subspace and noise subspace using orthogonal matrix vectorization techniques and later the noise subspace will be suppressed (Hyvarinen et al. 2001; Jensen and Hansen 2001). In model-based methods, trained models by GMM/HMM/CNN/DNN of clean speech and interference are used to segregate the speech. Model-based methods show satisfactory performance only when predefined models are available and the model of the speech is matched exactly with the predefined models (Dharmalingam et al. 2017; Gibak and Loizou 2010; Yi-nan et al. 2014; Naik et al. 2012; Naik and Ganesh 2012; Nilesh et al. 2013; Qazi et al. 2018; Sameti et al. 1998; Singhal et al. 2018; Trowitzsch 2017; Wang and Kun 2013; Yu et al. 2017; Zhang and Wang 2017). The remarkable single channel model-based system proposed by Nilesh et al. (2013) uses IBM (Ideal Binary Mask) and IWF (Ideal Wiener Filter) to segregate the target speech. The performance of the single channel model-based system is improved even at low SNR conditions. Another model-based system proposed by Gibak and Loizou (2010) uses Bayesian classifier to classify the noisy speech into a target and masker dominant T–F units. Dharmalingam et al. (2017) proposed a successful model-based system which uses SVM (support vector machine) model to segregate the target speech from the noisy mixture. All these approaches need some form of clean speech and noise but this may not be possible in real scenarios. Alternatively, a feature-based approach has been used for speech segregation since it does not need a prior information of the speech or noise. After Auditory Scene Analysis (ASA) has been proposed and, many researchers started to use ASA concepts to build feature-based speech segregation system are known as Computational Auditory Scene Analysis (CASA) systems (Brown and Cooke 1994; Brown and Wang 2005; Donald and Wang 2017; Harish and Rajavel 2014; Hu and Wang 2006).

A typical CASA system consists of auditory filtering, feature extraction, segmentation, grouping and re-synthesis. In auditory filtering, the acoustic mixture is decomposed into many T–F units using bandpass filtering and windowing technique. In feature extraction, auditory cues such as correlogram, cross-correlogram, pitch and onset/offset are extracted as features. In segmentation, each T–F units are segmented into speech T–F units and noise T–F units based on the extracted features. In grouping, the speech segments are grouped together and noise segments are grouped together to form speech group and noise group respectively. Each T–F unit in the speech group is denoted by ‘1’ and similarly each T–F unit in noise group is denoted by ‘0’ this constitute a binary mask. Finally, the obtained binary mask is used to synthesize the noisy speech into speech and noise in the re-synthesis stage. Various CASA systems have been proposed in the literature over the past three decades for monaural speech segregation (Brown and Cooke 1994; Brown and Wang 2005; Donald and Wang 2017; Harish and Rajavel 2014; Hu and Wang 2006). A CASA system proposed by Hu and Wang (2004) uses auto-correlation, cross-correlation, pitch tracking and amplitude modulation to separate the monaural voiced speech from the noisy speech signal. Wang (2012) proposed another CASA system namely tandem algorithm to separate the monaural voiced speech from the noisy speech mixture and showed significant improvement in the speech quality. The above two CASA systems concentrate mainly on separating the voiced speech from the monaural noisy speech mixture and do not deal with unvoiced speech. A CASA system proposed Hu and Wang (2007) segment the speech into voiced and unvoiced speech segments using onset and offset analysis. Later, Hu and Wang (2013) proposed an unsupervised model-based CASA system for co-channel speech segregation in which tandem algorithm and clustering technique are used to produce voiced binary mask. Onset and offset based segmentation scheme is used to generate the segments for unvoiced speech. Later, the voiced T–F binary mask and unvoiced segments are grouped together and an overall T–F binary mask has been generated for co-channel speech segregation. Wang et al. (2013) proposed a CASA based monaural speech segregation system to segregate both voiced and unvoiced speech from the noisy speech mixture. This system uses auditory cues such as correlogram, cross-correlogram, response energy for voiced speech and onset/offset for unvoiced speech to construct a binary mask. The author proposed a morphological image processing technique to smooth and retrieve the binary mask. The obtained binary mask contains some broken auditory fragments which will degrade the quality of the resynthesized speech.

This research work proposes a new method to compute the T–F binary mask of unvoiced speech and voiced speech and then fuse them effectively using Genetic Algorithm (GA). The systems proposed by Hu and Wang (2013) and Wang et al. (2013) use onset and offset analysis to obtain the unvoiced segments and merged with the voiced T–F binary mask. The proposed system uses the same onset and offset analysis to extract the voiced and unvoiced segments and then mapped into its equivalent T–F binary mask. Another T–F binary mask is obtained from the voiced speech binary mask computation via feature extraction and initial segmentation. The two T–F binary masks representing voiced and unvoiced speech are fused using a Genetic Algorithm (GA). In the fusion process, a Genetic Algorithm is used to appropriately weight the voiced and unvoiced T–F binary mask such that the quality and intelligibility of the speech is increased. The proposed system is evaluated using Cooke speech database (Cooke 1993) which consist of 10 speech utterances and 17 noises (Wang et al. 2013). The proposed system performance is measured in terms of Signal to Noise Ratio (SNR), Perceptual Evaluation of Speech Quality (PESQ) as a quality measure, Percentage of noise residue (\(P_{nr}\)) and Percentage of energy loss (\(P_{el}\)) as complementary error measure and coherence speech intelligibility index (CSII), normalized covariance metric (NCM) and short-time objective intelligibility (STOI) as the intelligibility measure (Hu and Loizou 2007; Ma et al. 2009; Taal et al. 2011). The outcome of this research work enhances the speech quality by improving the SNR and PESQ and improves the speech intelligibility in terms of CSII, NCM and STOI with respect to other existing systems.

The rest of the research article is prepared as follows. Section 2 describes the computation of voiced T–F binary mask, a new scheme of computing T–F binary mask for unvoiced speech, and finally the proposed GA based T–F binary mask fusion. In Sect. 3, the quality, intelligibility and complementary error measures are evaluated and their results are reported. Finally, Sect. 4 concludes the paper with the possible future extension of this research work.

2 Speech segregation system

The main contribution of this research work is to effectively fuse the T–F binary mask of voiced and unvoiced speech using the Genetic Algorithm such that the quality and intelligibility of the speech is improved. The proposed T–F binary mask fusion using a Genetic Algorithm is illustrated in Fig. 1 contains auditory filter bank, voiced binary mask computation, unvoiced binary mask computation; Genetic Algorithm based T–F binary mask fusion and speech re-construction of the signal.

Fig. 1
figure 1

Proposed GA speech segregation system

Full size image

2.1 Auditory filter bank

The proposed speech segregation system illustrated in Fig. 1 contains an auditory filter bank as the first element where the noisy speech mixture is processed through 128 channel gammatone filterbanks centered from 50 Hz to 8 KHz and its response is given in (Meddis 1988; Patterson et al. 1988) as

$$g_{c}(f_{c},t) = {\left\{ \begin{array}{ll} b_{w}^Nt^{N^{-1}}e^{2\pi b_{w}t}\cos (2\pi f_{c}t), \qquad \text {if }t\ge 0,\\ 0, \qquad \text {otherwise.} \end{array}\right. }$$
(1)

where \(b_{w}\) determines the bandwidth, \(f_{c}\) is the center frequency and N is the order of the filter. The response from the gammatone filter produces a two-dimensional Time–Frequency (T–F) mask called cochleagram.

2.2 Voiced binary mask computation

The main idea of this research work is to employ a feature extraction process to obtain the T–F binary mask for voiced and unvoiced speech. The schematic block diagram to obtain the voiced T–F binary mask is shown in Fig. 2. It consists of feature extraction, initial segmentation, pitch estimation, segments grouping and unit labeling. The following subsections briefly explain in detail of the above-mentioned components of voiced binary mask computation.

Fig. 2
figure 2

Voiced binary mask computation

Full size image

2.2.1 Feature extraction

Feature extraction is one of the important stages in the voiced binary mask computation where the response of the filter is given as an input to extract speech features such as energy, correlogram and cross-correlogram (Hu and Wang 2004). The reliable extracted speech features are used in the initial segmentation stage to segment each T–F unit into voiced dominant or noise dominant T–F unit.

2.2.2 Initial segmentation

The initial segmentation is the intermediate stage in voiced binary mask computation in which every individual T–F unit is segmented into voiced dominant or noise dominant T–F unit. According to Hu–Wang (2013), if the energy \(E(F_{c},T_{m})\) of a T–F unit is greater than a constant threshold \(\theta ^{2}_{H}\) (where \(\theta ^{2}_{H} = 50\)) the corresponding T–F unit is represented as voiced dominant and denoted as ‘1’ and the remaining T–F unit is represented as noise dominant and denoted as ‘0’. Likewise, the cross-correlogram of a T–F unit is greater than threshold \(\theta _{c}\) [0.985 as in Wang and Kun (2013)], the corresponding T–F unit is denoted as ‘1’ otherwise denoted as ‘0’. This approach reduces the quality of speech by missing some T–F units where voiced speech is less dominant than noise. It is known, the energy and cross-correlogram are the reliable features to segregate the voiced speech from the noisy speech mixture. By combining the energy and the cross-correlogram, a T–F binary mask \(M(F_{c},T_{m})\) for voiced speech is created as in Shoba and Rajavel (2017a) as

$$\begin{aligned} M(F_{c},T_{m}) = {\left\{ \begin{array}{ll} 1 \quad if \ \ E(F_{c},T_{m})\ge \theta _{AT}{(F_{c})} \& \ C(F_{c},T_{m})\ge \theta _{c},\\ 0 \quad {otherwise,} \end{array}\right. } \end{aligned}$$
(2)

The binary mask \(M(C_{H},f_{R})\) well represent the voiced segment of speech and failed to represent the unvoiced segment of speech.The above T–F binary mask \(M(F_{c},T_{m})\) well represents the voiced speech and denoted as ‘1’ and remaining T–F units are denoted as ‘0’.

2.2.3 Segments grouping and Unit Labeling

Segments grouping and unit labeling is the last stage in the voiced binary mask computation. The above stage captures the majority of the voiced speech T–F units and misses few of the T–F units where noise is more dominant. The missed T–F units are recovered by the process called as pitch estimation and are done by the summary of the correlogram of T–F units over all channels in Hu and Wang (2004) as

$$\begin{aligned} \S _{F}{(T_{m},\tau )} = \sum _{c}A_{H}(F_{c},T_{m},\tau ) \end{aligned}$$
(3)

The pitch estimation proposed in Hu and Wang (2004) estimates the pitch period at frame \(T_{m}\) in plausible pitch range of 2 to 12.5ms. By this estimation, most of the voiced speech T–F units are captured accurately by eliminating the noise components of the speech. The voiced speech T–F units captured from the pitch estimation is grouped and labeled as in Hu and Wang (2004) as

$$\begin{aligned} M_{v}{(F_{c},T_{m})} = {\left\{ \begin{array}{ll} 1 & \quad if \frac{A_{H}(F_{c},T_{m},\tau _{s}(T_{m}))}{A_{H}(F_{c},T_{m},\tau _{p}{(F_{c},T_{m})})} > \theta _{T} \\ 0 & \quad otherwise, \end{array}\right. } \end{aligned}$$
(4)

where \(\theta _{T}\) is the threshold equal to 0.85 which is the most appropriate value taken from Wang and Brown (1999) for labeling the particular T–F unit that determines the target speech as voiced speech denoted as ‘1’ and remaining background T–F units as noise denoted as ‘0’. The voiced binary mask computation stages finally produce a voiced T–F binary mask \(M_{v}{(F_{c},T_{m})}\) that captures almost all the voiced speech T–F units and also eliminates the noise present in the noisy speech mixture.

2.3 Unvoiced binary mask computation

The previous section explains various steps involved in the computation of the T–F binary mask for voiced speech. This section explains how a T–F binary mask is acquired in this research work for unvoiced speech. The schematic block diagram to obtain the T–F binary mask for unvoiced speech is shown in Fig. 3. It consists of a smoothing process, onset/offset detection and matching and multiscale integration and mapping of segments into the T–F binary mask. The following subsections explain the computation of the T–F binary mask for unvoiced speech.

Fig. 3
figure 3

Unvoiced binary mask computation

Full size image

2.3.1 Onset/offset analysis

The first step in the unvoiced binary mask computation is the onset/offset analysis. The sudden increase in intensity is called onset and sudden decrease in intensity is called offset. This can be obtained by taking the first derivative of the intensity with respect to time and then finds the peaks and valleys of the derivative. Actual onset/offset cannot be determined due to intensity fluctuations produced by the peaks and valleys of the derivate. To reduce the intensity fluctuations, smoothing is done which smooth the intensity over time and frequency. After smoothing, onset/offset detection and matching detects the onset and offset in every filter channel and merge them into onset and offset fronts. Each onset /offset fronts are paired to form segments. Finally the multiscale integration form segments at different scales (Hu and Wang 2007). The following subsections explain in detail about the onset/offset analysis.

Smoothing The first stage in the onset and offset analysis is smoothing. Smoothing is a process where the fluctuations will be reduced to produce a better solution. The initial intensity \(\vartheta (c,t,0,0)\) is smoothed with a low pass filter \(h(s_{t})\) over time and then \(\vartheta (c,t,0,s_{t})\) is smoothed with a Guassian kernel \(g(0,s_{c})\) over frequency. The response of the smoothing is given as in Hu and Wang (2007) as

$$\begin{aligned}&\vartheta (c,t,0,s_{t})=\vartheta (c,t,0,0)*h(s_{t}), \end{aligned}$$
(5)
$$\begin{aligned}&\vartheta (c,t,s_{c},s_{t} )=\vartheta (c,t,0,s_{t} )*g(0,s_{c}), \end{aligned}$$
(6)

where * represents convolution, the smoothing degree is denoted by \((s_{c}, s_{t})\). At the end of the smoothing process, the intensity fluctuations are reduced considerably but at the same time, major intensity changes of onset and offset are preserved.

Onset and offset detection and matching The next step after smoothing is the onset and offset detection and matching, which detects the onset and offset by finding the peaks and valleys of the derivative of the intensity as given in Hu and Wang (2007) as

$$\begin{aligned} \frac{d}{dt}{\vartheta (c,t,s_{c},s_{t})}=\vartheta (c,t,0,0)*\left[ \frac{d}{dt} h(s_{t})\right] *g(0,s_{c}), \end{aligned}$$
(7)

The peak and valleys are determined by setting a threshold value \(\theta _{T}=\mu +\sigma\) where \(\mu\) denotes mean and \(\sigma\) denotes standard deviation. A true onset is found by its corresponding peak higher than threshold \(\theta _{T}\) and offset is found between two onset peaks. In this way, onset and offset are determined for each filter channels. The common onset and offset found in adjacent channels are merged to form onset and offset fronts. Then the individual onset and offset fronts are matched to form segments (Hu and Wang 2007).

Multiscale integration The final step in the onset and offset analysis is the multiscale integration. The segments generated in the above step captures the voiced and unvoiced speech segments roughly but may miss some small voiced and unvoiced speech segments. To generate the segments accurately different scales are integrated from coarser to finer scale (Hu and Wang 2007). The coarser scale captures a majority of the segments containing speech events but misses some of the small segments. Segments locate more onset and offset positions at finer scale but do not form any new speech T–F segments. The finer scale captures more T–F speech segments and also more segments from the interference. Thus this process identifies voiced, unvoiced and some noisy speech segments as denoted as \(M_{v/uv}{(F_{c},T_{m})}\) which will be an important cue for speech segregation.

2.3.2 Mapping of segments into an unvoiced T–F binary mask

One of the proposals of this research work is to obtain the T–F binary mask for the unvoiced speech from segments \(M_{v/uv}{(F_{c},T_{m})}\) generated through the onset and offset analysis. Each segment of onset and offset analysis contains a unique label. The steps involved in the computation of the T–F binary mask for an unvoiced speech from these segments are as follows


Proposed algorithm

Step 1: :

Replace all the non-zero label value in the segments produced by the onset and offset analysis by binary 1.

Step 2: :

Next, compare the segment map of voiced speech \(M_{v}(F_{c},T_{m})\) obtained in Sect. 2.2 with the segment map of unvoiced and voiced speech obtained using onset and offset analysis \(M_{v/uv} (F_{c},T_{m})\).

Step 3: :

Then, identify the overlapping region of \(M_{v}(F_{c},T_{m})\) and \(M_{v/uv} (F_{c},T_{m})\) and complement all the T–F units of \(M_{v/uv} (F_{c},T_{m})\) within that overlap region. This process will remove all the voiced segments of \(M_{v/uv} (F_{c},T_{m})\) as follows

$$\begin{aligned}&M_{vr/I}(F_{c},T_{m})= M_{v/uv}(F_{c},T_{m}) - M_{v}(F_{c},T_{m}) \quad and\nonumber \\&M_{uv/I} (F_{c},T_{m})= 0 \quad if \, M_{vr/I}(F_{c},T_{m})< 1, \end{aligned}$$
(8)
Step 4: :

The resultant binary matrix \(M_{uv/I}(F_{c},T_{m})\) now contains only the unvoiced speech segments with a little bit of interference. To further remove the interferences, the energy of the T–F frame corresponding to the binary 1 in \(M_{uv/I}(F_{c},T_{m})\) is computed via correlogram similar in (Hu and Wang 2004) as follows

$$\begin{aligned}&A_{uv}(F_{c},T_{m},\tau ) \nonumber \\&\quad = \frac{1}{N_c} \sum _{n=0}^{N_{c}-1}h_{uv}(F_{c},T_{m}T-n)*h_{uv}(F_{c},T_{m}T-n-\tau ), \end{aligned}$$
(9)

where \(h_{uv}\) is the hair cell output corresponding to binary 1 in \(M_{uv/I}(F_{c},T_{m})\) and * denotes convolution.

Step 5: :

If the energy of the T–F units corresponding to the binary 1 in \(M_{uv/I} (F_{c},T_{m})\) is greater than a threshold \(\theta _{A}\) those T–F units are marked as 1 to represent it as unvoiced speech segments otherwise it is zero as follows

$$\begin{aligned} M_{uv}{(F_{c},T_{m})} = {\left\{ \begin{array}{ll} 1 \ \ if \ \ A_{uv}(F_{c},T_{m},\tau ) > \theta _{A} \ \ \\ 0 \ \ \ otherwise, \end{array}\right. } \end{aligned}$$
(10)

The value for \(\theta _{A}\) is determined via experimentation and the resultant matrix \(M_{uv}(F_{c},T_{m})\) represent the T–F binary mask for unvoiced speech.

2.4 Proposed GA based T–F binary mask fusion

Another proposal of this research work is to effectively fuse the T–F binary mask representing the voiced and unvoiced speech such that the speech is enhanced using a Genetic Algorithm. The Genetic Algorithm (GA) is an optimization and search technique based on the principles of genetics and random selection. An initial population of individuals (weights for fusion) is made and by iterative use of the genetic operators (selection, crossover, mutation) an optimal weight is reached according to the defined fitness function (Rajavel and Sathidevi 2012, 2011). The GA used in this work to obtain the correct weight factor \(\alpha\) to fuse the T–F binary mask of voiced and unvoiced speech and in turn improves the speech quality and intelligibility according to the following equation as

$$\begin{aligned} Opt_{{BM}_{T--F}}(F_{c},T_{m})& = \alpha *M_{v}(F_{c},T_{m}) \\&\quad +(1-\alpha )*M_{uv}(F_{c},T_{m}) \quad for \quad 0 \le \alpha \le 1, \end{aligned}$$
(11)

Initially, \(\alpha\) has chosen as 0.5 to give equal weights to voiced and unvoiced T–F binary mask and conducted the experiment to separate the estimated speech from the noisy speech mixture with various noises. The experimental result does not show any significant improvement in speech as compared to the existing systems (Shoba and Rajavel 2017b, 2018a, b; Wang et al. 2013). It is known, the low-frequency filter channel may represent the voiced speech more dominantly than high-frequency filter channel and vice versa for the unvoiced speech. Hence there is a need for a scheme to give adaptive weights to the T–F units in the filter channel which in turn change the Eq. (11) as

$$\begin{aligned}&Opt_{{BM}_{T--F}}(F_{c},T_{m}) \nonumber \\&\quad = \alpha *M_{v}(F_{c},T_{m})+(1-\alpha )*M_{uv}(F_{c},T_{m}) \quad for\nonumber \\&\qquad 0 \le \alpha \le 1 \quad for \, 1 \ldots N_{c} \quad and\nonumber \\&Opt_{{BM}_{T--F}}(F_{c},T_{m}) \nonumber \\&\quad = \alpha *M_{uv}(F_{c},T_{m})+(1-\alpha )*M_{v}(F_{c},T_{m}) \quad for\nonumber \\&\qquad 0 \le \alpha \le 1 \quad for \, N_{c}+1 \ldots N_{T}, \end{aligned}$$
(12)

where \(N_{c}\) is the number of filter channel where voiced speech dominates which is determined via experimentation and \(N_{T}\) is the total number of filter channels in \(Opt_{{BM}_{T--F}}(F_{c},T_{m})\). The proposed Genetic Algorithm (GA) based scheme to effectively fuse the T–F units of voiced and unvoiced T–F binary mask is as follows:

figure a

Finally, the above proposed algorithm returns the optimal weight factor \(\alpha\) which produce the speech with good quality and intelligibility.

Fig. 4
figure 4

Spectrogram results of the (a) clean speech - “ I ll willingly marry Marilyn” (b) noisy signal with white noise at 0 dB SNR (c) enhanced speech using IASG system (d) the proposed GA based fusion scheme

Full size image

2.5 Reconstruction of the speech signal

The final stage in the proposed research is to reconstruct the speech signal from the T–F binary mask obtained using a Genetic Algorithm (GA) based fusion scheme. The re-construction of the speech signal used in this work follows the same procedure proposed by Weintraub (1985). Figure 4 represents the spectrogram of the (a) Clean speech signal “ I ll willingly marry Marilyn”. (b) Noisy speech mixture (White noise) at 0dB SNR (c) Enhanced speech signal of using IASG system. (d) Enhanced speech using the proposed Genetic Algorithm (GA) based fusion scheme. From the observation, the final segregated speech from the proposed algorithm contains both voiced and unvoiced speech.

Fig. 5
figure 5

Description of noise

Full size image

3 Evaluation results and comparison

The evaluation of the proposed speech segregation system is done using 10 speech utterances from Cooke (1993) database and 17 noises (Wang et al. 2013) are mixed at specific SNR ranges to yield 170 (\(17\,\times \,10\,\times \,1\)) noisy speech utterances and at different SNR ranges (\(-\,5, \,0, \,5,\,10\,{\rm{and}}\,15~\mathrm{{dB}}\)) to yield 835 (\(17\,\times \,10\,\times \,5\)) noisy speech utterances. Figure 5 shows the noises used for this research work and the proposed system is evaluated with quality measures such as Signal-to-noise-ratio (SNR) and Perceptual Evaluation of Speech Quality (PESQ) and intelligibility measures such as short-time objective intelligibility (STOI), coherence speech intelligibility index (CSII), and normalized covariance metric (NCM) (Hu and Loizou 2007; Ma et al. 2009; Taal et al. 2011). The performance of the proposed system is also measured using two error measures such as Percentage of energy loss (\(P_{el}\)), Percentage of noise residue (\(P_{nr}\)). The proposed system is measured in terms of both quality and intelligibility. Since the existing (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004) evaluation results might be an appropriate one for comparing the quality measure with the proposed system and may not be an appropriate one for comparing intelligibility measure with the proposed system. Hence, these two systems have been re-evaluated using the Cooke Speech database and 17 noise and then the results (quality and intelligibility) are compared with the proposed system. The proposed work is implemented using a personal computer with an Intel ®  Core\(^{\mathrm{TM}}\) (i5-3230M) 2.60 GHz processor, 8 GB RAM. The software used is MATLAB (R2016b) installed in Windows 10 64 bit operating system.

Fig. 6
figure 6

Proposed system performance with the other systems (SNR(dB)-Quality)

Full size image

The SNR is a direct criterion to measure the quality of speech and is determined as in Wang et al. (2013)

$$\begin{aligned} SNR = 10\log \left( \frac{\sum _{n}{S_{IN}(n)^2}}{\sum _{n}{(S_{IN}(n)-S_{OUT}{(n)})^2}}\right) , \end{aligned}$$
(13)

where the original speech is named as \(S_{IN}(n)\) and the segregated speech is named as \(S_{OUT}{(n)}\). Figure 6 depicts that there is an improvement in SNR over all noises when compared with the existing system. The proposed system improves the SNR with an average value of 10.74 dB with input mixture, 3.47 dB with (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004), 2.19 dB with (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and 0.83 dB with (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b) respectively. From the observation, the proposed system shows a high SNR improvement for noises \(N_{3}\), \(N_{7}\), \(N_{10}\), \(N_{14}\) as it retains most of the speech information and there is a modest SNR improvement for noises \(N_{11}\), \(N_{15}\), \(N_{16}\), \(N_{17}\) due to significant spectral overlap between speech and noise. The proposed algorithm merges the unvoiced and voiced speech effectively using a Genetic Algorithm and in turn increases the quality of the speech when compared with the other systems in terms of SNR measure (Hu and Wang 2004; Wang et al. 2013).

Fig. 7
figure 7

Proposed system performance with the other systems (PESQ-Quality)

Full size image

An objective measure called perceptual evaluation of speech quality (PESQ) (Pichevar and Rouat 2005) is used to measure the segregated quality of speech for the proposed system. The bigger the PESQ worth suggests a greater performance approach. Figure 7 depicts the proposed system performance in comparison with other systems (Wang et al. 2013). The PESQ improvement for the proposed system increases with an average value of 0.4 with (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004), 0.2 with (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and 0.08 with (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b) respectively. From the comparison, the PESQ results are superior for all types of noises and produce good quality of speech.

The above two common criterion SNR and PESQ relate to the speech quality measure and does not provide more information approximately how the segregated target speech is different from that of the clean speech signal. To do so, two complementary error measures \(P_{el}\) and \(P_{nr}\) are evaluated for the proposed speech segregation system to achieve low errors.

The percentage of energy loss \(P_{el}\) is measured by the quantity of target speech excluded from the segregated speech and determined by

$$\begin{aligned} P_{el} = \frac{\sum _{n}{e_{{s}_{1}}^2(n)}}{\sum _{n}{I_{s}^2(n)}}, \end{aligned}$$
(14)

The percentage of noise residue \(P_{nr}\) is measured by the quantity of noise included in the segregated speech and determined by

$$\begin{aligned} P_{nr} = \frac{\sum _{n}{e_{{s}_{2}}^2(n)}}{\sum _{n}{O_{s}^2(n)}}, \end{aligned}$$
(15)
Table 1 Error measure - \(P_{el}\) and \(P_{nr}\) performance results of the proposed system and the other systems
Full size table

\(I_{s}(n)\) is the re-synthesized speech from IBM (Ideal Binary Mask). \(O_{s}(n)\) is the re-synthesized speech from the proposed system.

\(e_{s1}\) is the signal energy present in \(I_{s}(n)\) but missing in \(O_{s}(n)\).

\(e_{s2}\) is the signal energy present in \(O_{s}(n)\) but missing in \(I_{s}(n)\).

The \(P_{el}\) and \(P_{nr}\) of the proposed system are evaluated and compared with other existing systems (Hu and Wang 2004; Wang et al. 2013) as shown in Table 1. Table 1 depicts that the energy loss is very less for all types of noises and the average value of energy loss \(P_{el}\) is 2.21% and the average value of noise residue \(P_{nr}\) is 2.23% which is less than (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004), (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b). The proposed system of \(P_{nr}\) retains 97.77% ((100-\(P_{nr}\))%) of the original speech information.

Table 2 The STOI (Intelligibility) of the proposed system and other systems at specific SNR range
Full size table
Fig. 8
figure 8

The performance of the proposed system and other system (STOI) intelligibility measures at different SNR ranges

Full size image

The STOI intelligibility measure shown in Table 2 observed that the average STOI value for the proposed system shows improvement [(i.e) 0.91 for the proposed system, 0.84 for (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b), 0.79 for (Wang–Yu) Morphological(MIPS) system (Wang et al. 2013) and 0.74 for (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004)] when compared with other existing system. Also observed that, there is a very low improvement in STOI particularly for noises such as \(N_{11}\), \(N_{16}\), \(N_{17}\) since it contains noise components present in the target speech and there is a decent improvement in STOI for noises such as \(N_{4}\), \(N_{7}\), \(N_{10}\) and \(N_{14}\) since most of the target speech is segregated accurately by eliminating the interference completely. Another experiment has been conducted with different SNRs ranging from \(-\,5,\, 0,\,5,\, 10 \mathrm{{ and}}\,15\,\mathrm{{dB}}\) for the proposed system and the other existing system for noise \(N_{1}\) as shown in Fig 8. From the graphical representation of the Fig. 8, it is observed that the proposed system shows a gradual improvement in STOI for low SNR ranges such as -5dB, 0dB and 5dB and there is a very low improvement in STOI for high SNR ranges. The observation of Table 2 and Fig 8 concludes that the intelligibility (STOI) of the speech is increased even at low SNR ranges.

Table 3 The CSII (Intelligibility) of the proposed system and other systems at specific SNR range
Full size table
Fig. 9
figure 9

The performance of the proposed system and other system (CSII) intelligibility measures at different SNR ranges

Full size image

The proposed system is evaluated using another measure (intelligibililty) called CSII as shown in Table 3. It is observed from Table 3 the average CSII value for the proposed system shows improvement ((i.e) 0.89 for the proposed system, 0.83 for (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b), 0.75 for (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and 0.67 for (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004)) when compared with other existing system. The keen observation of Table 3 depicts that the target speech components are segregated accurately from the noisy mixture for noises such as \(N_{5}\), \(N_{6}\) and \(N_{10}\) as compared with other noises. The performance of the proposed system and other existing system for noise \(N_{1}\) is measured in a different view by conducting experiments at different SNR ranges such as -5 to 15dB in steps of 5dB as shown in Fig 9. The Fig 9 of the proposed system shows a gradual improvement over SNR ranges of -5dB, 0dB and 5dB and if SNR range increases 5dB, the improvement of CSII value is low because it fails to suppress the noise-dominant units at high SNR levels.

Table 4 The NCM (Intelligibility) of the proposed system and other systems at specific SNR range
Full size table
Fig. 10
figure 10

The performance of the proposed system and other system (NCM) intelligibility measures at different SNR ranges

Full size image

Apart from intelligibility measures such as STOI and CSII, another intelligibility measure called NCM is evaluated as shown in Table 4. The performance of the NCM measure in Table 4 shows an average value of 0.82 for the proposed system, 0.76 for (Shoba–Rajavel) Image processing (IASG) system (Shoba and Rajavel 2017b), 0.68 for (Wang–Yu) Morphological (MIPS) system (Wang et al. 2013) and 0.61 for (Hu–Wang) pitch tracking (PT) system (Hu and Wang 2004) respectively. Similar to Figs. 8 and 9, the NCM measure has been evaluated at different SNR ranges from \(-\,5\) to 15 dB. Figure 10 depicts that there is a significant improvement in NCM value for low SNR ranges when compared with high SNR ranges. From the experimental results, it is proved that the unvoiced and voiced speech T–F units are segregated effectively by the proposed Genetic Algorithm (GA) based fusion and in turn improves both the quality and intelligibility of the speech.

4 Conclusion

This research work proposed a Genetic Algorithm (GA) based fusion scheme to segregate the target speech from the noisy speech mixture. The proposed work computes the T–F binary mask representing the unvoiced speech and effectively fused this binary mask with another mask representing the voiced speech using a Genetic Algorithm. The Genetic Algorithm (GA) based fusion enhances the T–F binary mask by representing both voiced and unvoiced speech and in turn, shows significant improvement in speech quality and intelligibility. The experimental results of the proposed system show improvement in SNR and PESQ by increasing the quality of speech and considerably reduce the noise residue and energy loss by deteriorating the \(P_{el}\) and \(P_{nr}\) value and also show a significant improvement in speech intelligibility as compared with the other existing systems. The Genetic Algorithm (GA) based fusion scheme gives the accurate optimal solution for fusing the voiced and unvoiced speech for a known set of speech and noise samples and may not be an optimal one for unknown speech and noise samples. The future course of this research will focus on this issue and propose a training based framework utilizing a neural network to deliver the optimal solution for all speech samples.