Recognition Using DNN with Bacterial Foraging Optimization Using MFCC Coefficients

Speech is one of the most effective ways of communication. Speech recognition is done by human beings every day [1]. It means to be able to recognize a foreign language, and be able to pronounce the common sounds [2]. Speech recognition technology refers to a system capable of recognizing the “miracle” of conversational speech. This is so because it is difficult to construct a computer that can understand spoken discourse on any topic in any situation because of the difficulty of the process [3].

Speech recognition is the mechanism by which a machine recognizes spoken word. Signal processing converts input speech into identifiable speech form [4]. Speaker (or lecturers or teachers) recognition method can be applied by studying the voiced/unvoiced components or by evaluating the speech energy distribution [5]. Since the model is modified to best reflect observed expression, it is called model-based system. Speech recognition has to be conducted in various ambient environments, thus, the features derived have to be resilient to background noise and system mismatch [6]. Speech emotion recognition technology is also desirable for applications which require natural man-machine interaction such as web movies and computer tutorial programs where the response of these systems depends on the detected emotion [7].

Speaker recognition is essential to the field of communication and surveillance [8]. The recognition process was tested by matching training and test results. Several experiments were performed to develop this kind of recognition process. Some disturbing and distressing shortcomings of the approaches are linear channel distortion, reverberation, and additive noise [9]. Here, the feature extraction of audio is regarded as a difficult task due to the existence of non-stationary noise and reverberation [10]. The characteristics which are used by the speech recognition system are fundamental and spectrum frequency histograms, linear prediction cepstral coefficients (LPCC) [11], instantaneous spectra covariance matrix, averaged auto-correlation, and MFCC (Multi-Frequency Cepstral Coefficients). LPCC and MFCC has a major role in identifying speakers [12].

The best among the means of identification of speakers is MFCC [13]. The local spectral properties of the speech signal can be analyzed using MFCC. The learned features are measured for the full dataset samples and the learned features are saved for speaker recognition [14]. Both training and research set samples are taken into account in MASS [15].

This model also helps users to talk normally without having to pause. They use subword sounds as basic unit of recognition in speech recognition [16]. This is the sequence of production process which can produce bigger linguistic unit such as words/sentences thus recognition [17]. Since the number of subword units are small (typically 45 for Indian languages), collecting adequate data of subword is not difficult [18]. Adding a new word is easy because all that is required is a simple re-wording of an existing word.

Speech recognition, popularly also known as Automatic Speech Recognition (ASR) is the method of translating speech signal to a sequence of words. Sample speech is initially gained from the user in a controlled way. The proposed model processes voice signals and extracts discriminatory information from the speaker. The proposed model extracts feature from the input voice by dividing the voice samples into small segments. Today, we can conveniently store the databases of speech recognition for different purposes thanks to the development of computer technology. The complete design of the work is seen in Figure 1.

1.png

Figure 1. Speech recognition system model

In speech recognition positions, MFCC is outstanding among the best prototypes used by machine speech recognition and speaker recognition programs, and mesh store investigation is made. Logarithmic frequency bands are commonly used in speech recognition applications. Approximates the human auditory better than most strategies. In order to obtain the membership function coefficients from the filtered signal, remove discontinuities through the hamming window to get it correctly.

A DNN is a type of network with a fixed degree of intricacy and with dynamic boundaries. DNN utilizes a systematic technological process for managing data. A DNN network with abundant layers usually incorporates feature extraction and feature organization into a signal learning bank. These NN models have demonstrated eminent success in extensively recording contemporary designs. The layers used in DNN are represented as,

$L_{1}=T\left(N_{1} x\right)+\log (a W i)$  (1)

where, N₁ is weight of initial. Activation function determines whether a neuron will be activated or not. The activation function is designed to introduce non-linearity into the output of a neuron. The activation function determines what the output of the network will be. This is what an activation function of an ANN does too. It takes the output of the previous cell and transmits it into another cell which is its input. The logistics function special case can be defined as,

$F(x)=e^{-x^{2}}$  (2)

Here, x denotes the input of activation. The mapping between the current and next hidden layer after getting the first hidden layer is given as,

$a_{1}=F\left(W_{1} a_{l}+b_{1}\right), l=2, \ldots, L$  (3)

2.png

Figure 2. Hidden layer model

where, ‘al’ is all layers count considering BFO parameters and b is the parameters that are relevant for speech recognition. The hidden layer model using DNN is depicted in Figure 2.

Bacterial Foraging optimization algorithm

A BFO algorithm is based on three processes; Chemotaxis, Replication and Elimination Dispersion. BFO algorithm were discussed here.

BFO Algorithm

Input: Initialize Population and Parameters T,P, N_t, N_c, P(i) {i=1,2,3...⍬}.

Step-1: Perform Preprocessing as:

$e\left(\boldsymbol{V}_{i}^{r^{\prime}}\right)=p\left(\boldsymbol{V}_{i}^{r^{\prime}}\right)-\beta p\left(\boldsymbol{V}_{i}^{r^{\prime}}-1\right)$  (4)

$e\left(\boldsymbol{v}_{i}^{t^{\prime}}\right)=p\left(\boldsymbol{V}_{i}^{t^{\prime}}\right)-\beta p\left(\boldsymbol{V}_{i}^{t^{\prime}}-1\right)$  (5)

where, ‘e’ is thge processing function, v is the vector for parameter calculation, β is the threshold limit for processing parameters.

Step-2: Create a sample iteration with same properties.

Step-3: Chemotaxis: The process of chemotaxis is the center of the algorithm, simulating E's foraging behavior. Shifting and tumbling coli. Bacteria tumble more often in poorer regions, while bacteria travel more often in areas where food is more plentiful.

The chemotaxis activity of the ith bacterium can be defined as:

θ i (j + 1, k, l) = θ i (j, k, l) + C(i) ∗ dcti    (6)

dcti = ∆(i) ∆T(i)∆(i)   (7)

where, the θ i (j, k, l) represents the ith bacterium at the jth chemotactic, kth reproductive, and lth elimination–dispersal steps.

Step-4: For i=1,2..S.

1. Calculate fitness function FF as

FF(i,j,N(V))=  θ i + dcti      (8)

2. A random variable is a vector between −1 and 1.
3.  Tumble: generate a random vector D(i)ÎRⁿ with each element D_m(i), m= 1,2,….,S,a random number on [ -1,1].
4.  Move: let (2.1). This results in a step of size C(i) in the direction of the tumble for bacteria i.
5.  Compute J(I,j+1,k,l)with qⁱ(j+l,k, l).
6.  Swim:

Step-5: let m=0 (counter for swim length).

while m<N_s (if not climbed down too long).

1.  let m = m+1
2.  if J(I,j+1,k, l) < J_last,

let J_last = J (I,j+1,k, l), Then, another step of size C(i) in this same direction will be taken as (2.1) and use the new generated q^I (j+1,k, l) to compute the new J(I,j+1, k, l).

3. else let m = N_s.

Go to next bacterium (i+1): if I ¹ S go to (b) to process the next bacteria.

Step-6: If j<N_c, go to Step 3. In this case, continue chemotaxis since the life of the bacteria is not over.

Step-7. Reproduction.

For the given k and l,

for each I = 1,2,…..,S, let

$J_{\text {health }}^{i}=\sum_{j=1}^{N_{c}+1} J(i, j, l)$  (9)

The health of the bacteria. Sort bacterium in order of ascending values (J_health).