Speaker Identification Based on Physical Variation of Speech Signal

Speech communication is a simplest way of communication. In speech signal processing speech signal delivered the information about the message or text being spoken [1]. Speaker identification is a biometric sensation under speech signal processing application that identifies the speaker [2]. Speech signal not only delivers the information about the message but also about the speaker identity, speaker health, emotion, age, physical condition, gender information etc. So that speech signal has an independent role in research in the application of speech signal processing [3]. There is a dissimilar area of research in speech signal dispensation like voice recognition, broadcaster identification, and speech enhancement. Identification of speaker is to identify an unknown speaker depended on his voice or dialect [4]. Research in speaker identification is the wide area of application such that securities control, forensic casework, confidential information etc. [5]. In this study a method for the identification of textual speakers by physical variations of the voice are proposed [6]. The language ID based on text-related, necessary speaker says the ordinary phrase or message [7]. In considerable research for voice disguise by two methods one of them electronic disguised or electronic variation method second non-electronic disguised or physical variation method of voice [8]. In electronic disguised method voice disguised by electronically but in case of non-electronic disguised or physical variation method depends on the physical condition of speakers [9]. This proposed approach of identifying a speaker based on non-electronically disguised text-dependent or physical variation method of speech. In this work considered as speakers normal voice (NV) disguised by ten method fast speech (FS), slow speech (SS), raised pitch (RP), lower pitch (LP), mask on mouth (MM), bite off block (pencil) (BB), chew gum (CG), pinched nostril (PN), object in mouth (OM), foreign ascents (FA). For speaker identification, it is most important to extract the feature of speaker’s voice [10]. There is a different method of feature extraction like MFCC, linear predictive cepstral coefficients, vector quantization etc. In this proposed method used MFCC feature extraction technique [11]. In existing research, that is based on forensic automatic speaker recognition (FASR) technique [12]. Here proposed a method for speaker identification by auditory feature analysis in addition to its arithmetical mean values and correlation coefficients. Analysed the existing model and designed using feature extraction technique [13]. For speaker identification selected our proposed work 20 candidates ranging between 20 to 25 years age. They converse the general text or message during a regular influence as well as a physical variation of their voice [14]. There are two type of feature extractions technique first one MFCC as discussed in this manuscript second feature extraction technique are Linear predictive cepstral coefficients (LPCC) techniques. The LPCC performances degrade in noisy environment. So, MFCC is better than LPCC.

By using the same approach like normal voice take the voice sample of all twenty speakers by ten types of physical variation method [15]. For identification of speaker from its disguised voice or physical variation of voice is identical. By evaluating all these things, this task segmented into two stages training stage as well as the testing stage [16]. By this, all the speakers statistical mean and correlation coefficients distributed at this stage with specific details [17]. With all the acoustic features details instruction and test data approved from side to side the classifier algorithms that is dependent on statisticals correlation coefficients and mean of regular speech feature and physical variation of voice [18].

The feature exaction of speech signal is to amplified the spectrum where hearing becomes sensitive is processed in the front-end processing of speech signals. As a result, it only amplifies the high-frequency portion of the spectrum because the lower frequency section has sufficient energy. As a result, pre-emphasis is crucial in spectrum analysis [19]. It equalises the natural slope before the spectral research, boosting the accuracy of the analysis. Figure 1 (a, b) shown before and after pre-emphasis, a time domains speech representations of a speech sentences were shown. The pre-emphasis filters are used as high-pass filter of first order. The equations of the filter in time domains are expressed: the input signal s[n] as well as pre-emphasising coefficients. The range was taken 0.9 to 1.0, the speech signal is expressed in time domain equation.

Because the speech signals are continually changing, it is assumed that it does not change considerably on short time scales for the reason of simplicity. As a result, the speech signal is divided into the multiple frames of 20-40ms durations, as shown in Figure 2. If the frames are substantially smaller, enough sample is required to achieve a reliable spectrum estimate; nevertheless, if the frame is much larger than typical, the signal changes greatly across the frame.

Windowing takes advantage of the windows to make the signal equal to zeros at both from the starting to the last. It decreases signal discontinuities and spectrum distortion. Frame size and frame shift affect the speech produced by each window frame. The numbers of millisecond between the left ends of consecutives window are known as frame size, whereas the numbers of millisecond between the right end of consecutives windows is known as frame shift. The original signal s[n] is multiplied with the windows w[n] at time n to produce a resultant signal.

Recognizing speakers involves being able to receive a voice signal, recognise a speaker in a voice signal, and then recognise the speaker again. Waveform transformation is achieved by using a parametric representation to reduce the data rate while increasing the amount of processing and analysis that can be done. Good classification, in turn, comes from exceptional qualities [20]. The MFCC, a method of speech extraction, was described in this chapter. It has been shown in several contexts that these methods work, and are both effective and practical [21]. Previously described processes have been tweaked by researchers to increase their tolerance for noise, enhance their strength, and reduce the amount of time they take. For another thing, the issue of where to apply the technology is not settled [22].

1a.png

1b.png

Figure 1. Speech sample waveforms a). Prior to pre-emphasis b). Post pre-emphasis [19]

2.png

Figure 2. Framing of speech signal [19]

For Statistical analysis, first pre-emphasized technique used as a high-pass filter. Short-term speech features are represented by MFCCs, which are taken from their spectrum. It's frequently utilised in acoustic signal analysis. The MFCCs' scheme is well-known and well approved. It describes the relationship between the humans ear critical bandwidth as well as the Mel-frequency scale, in which filters are placed logarithmically above 1000 Hz and linearly below 1000 Hz. To eliminate discontinuities, The signals is first segmented into frame, and then each frame is multiplied by a Hamming window. The DFT of each frame are then calculates, and the log of the amplitude spectrums are measured. Finally, the speech spectra are smoothed using the discrete cosine transform (DCT), which aids in the generation of cepstral feature vectors for each frame. An analogous speech waveform is first converted to digital speech using MFCC. Its sampling rate is larger than or equivalent to 10000 Hz to reduce aliasing problems in speech.

The process of MFCC generation is shown in Figure 5. The contribution voice signal v[n] as well as its postponement translation of voice gesture v[n-1] and used µ pre-emphasized coefficient. This Pre-emphasized coefficient be used range from 0.9 to 1.0, after this the filter equation can be articulated the same as:

$y[n]=v[n]-\mu v[n-1]$                  (1)

By using the Hamming windowing techniques for dropping the supernatural dissimilarities and alteration of the voice signal and continue the non-interruptions of the voice gesture of every enclose, take 16ms-20ms where v[n] be the voice gesture and w[n] is function of window [1, 15].

$p[n]=v[n] * w[n]$                        (2)

where, w[n] is the window functions.

$w[n]=\left\{\begin{array}{lr}0.54-0.46 \cos \left(\frac{2 \pi n}{L}\right) & 0<L<n-1 \\ 0 & \text { otherwise }\end{array}\right.$             (3)

5.png

Figure 5. MFCC feature extraction

Frequency domain analysis of the voice signal after windowing:

$P_{i}(k)=\sum_{n=1}^{N} p_{i}(n) e^{-j 2 \pi k n / N} \quad 1 \leq k \leq N$                      (4)

Statistical moments extract acoustic vectors with similar durations as the number of MFCC characteristics changes with time. Here, x(n) is an N-frame speech signal whose MFCC vectors is vij, where j denotes the feature coefficients and it’s denoted by the frame numbers, and it may be represented as:

$V_{j}=\{v 1 j, v 2 j \ldots, v N j\};$               (5)

where $j=1,2 \ldots \ldots L$                      (6)

In this manuscript different two methods of statistical coefficients are derived. First the means Ej of each MFCC’s features coefficient of Vj is extracted. It is derived the correlation coefficients $C R_{j j^{\prime}}$ of the different MFCCs features of $V_{j}$ and $V_{j}$. The procedure is described in Eq. (7) and Eq. (8) respectively.

$E_{j}=E\left(V_{i}\right) ; j=1,2, \ldots \ldots L$               (7)

$C R_{j j^{\prime}}=\frac{\operatorname{cov}\left(V_{j}, V_{j^{\prime}}\right)}{\sqrt{\operatorname{var}\left(V_{j}\right)} \sqrt{v a r\left(V_{j^{\prime}}\right)}} ; 1 \leq j<j$              (8)

Similarly, the mathematical moment of the delta MFCC $\left(K_{\triangle M F C C}\right)$ and $\triangle \triangle \mathrm{MFCC}\left(K_{\triangle \triangle M F C C} \quad \right)$ are determined. At last, by combination of $K_{M F C C}, K_{\triangle M F C C}$   and $K_{\triangle \triangle M F C C}$   a statistical moments 'K' is created, which is given as:

$K=\left[K_{M F C C}, K_{\triangle M F C C}, K_{\triangle \triangle M F C C} \quad \right]$                (9)

The most important classifier used for statically classification technique. The speaker’s classification work can be categorized addicted to two types: test-dependent and training-dependent. Altered type of a classifier that be being used in the analysis are SVM classifier DT, NB, LDA and LR. The main benefit of these cepstral coefficients are derived from MFCC tend to be uncorrelated, making the systems easier to understand. Each frames MFCC features solely represent the power spectral envelope of the following frame. Derivative coefficients are used to extract speech-specific information in the dynamics. Each delta feature represents variance between frame, where each double delta feature represents variations between frame in the delta feature characteristics that follow.