Experimental Speech recognition from pathological voices

HADJI SALAH

Laboratory of nano-materials (LANSER) of Energy Center (CRTEn)

Technopole of Borj-cedria, Hamm-lif, Tunis, 2050

TUNISIA

Abstract: Speech recognition has been the subject of quite a few research subjects as it is the adequate

means for dynamic, efficient and interaction Human communication simultaneously using the two

phenomena of phonation and hearing between speakers, the applications of its searches are enormous

MFCC, PLP, LPC, RASTA and other in the form of combination (hybridization) namely: PLP

RASTA, MFCC PLP etc. They are used in data transmission, speaker recognition and even in speech

synthesis.

Keywords: speech signal, Parametrisation, SVM, pathological voices, classifier, MFCC, PLP,

RASTA, LPC

Received: April 18, 2021. Revised: August 17, 2022. Accepted: September 23, 2022. Published: October 31, 2022.

1. Introduction

The extraction of acoustic parameters or

recognition of a speech signal, its role is to

extract from a speech signal the most relevant

information possible in order to be able to

make a separation between the sounds [8]. The

extracted information is presented as a

sequence of acoustic vectors. In order to be

able to extract these parameters, several

methods exist, taking into account the

superposition of the noises of the sounds, we

will make a comparison of the different

this paper we will give a brief overview on the

signal processing tools such as short-term

energy and weighting windows,then see the

different speech signal parameterization

methods which are: LPC (Linear Predictive

Coding) analysis, Homomorphic or Cepstral

analysis on which the MFCC (Mel Frequency

Cepstral Coefficient) is based, PLP (Predictive

Linear Perceptual) and PLP-RASTA (Real

Ative SpecTrA).this involves using an SVM

used samples of corpus from pathological

signals (Nodule and Oedema) and another

from normal signals to make this classification,

(see Fig.) Shows us the principle of

classification.

classification, we used a One VS all type of

algorithm, that is to say "one against all", this

algorithm consists of taking a signal and

comparing it with all the classes.

2. Parameterization methods

There are several methods of parameterization;

there are those, which are based on the

perception of the human ear like the MFCC,

PLP and others, which are interested in the

Obtaining the Cepstral Coefficients at the Mel

scale was developed in 1980 by Davis and

Mermelstein, to do this it is necessary to apply

a Hamming window to each frame of the

signal, we then obtain a Cepstral characteristic

vector per frame, then we apply the Discrete

Fourier Transform (DFT). let us then keep the

Logarithm of the amplitude spectrum, then

after smoothing the spectrum let us apply the

Discrete Cosine transform to have the Cepstral

coefficients (see figure).

This step in the process is to emphasize the

high frequencies, this result in increasing the

energy at the higher frequencies.

Y (n) =X[n]-a. X [n-1] (1)

Stage 2: Segmentation into frames: this stage

consists in fragmenting the signal into frames

of 20 to 40 ms. The speech signal is split into

N samples. Adjacent samples are spaced by M

(M<N), typically the values used are M=100

and N=256.

such that:

N: number of samples in each frame

Y[n]: Output signal

X[n]: Input signal

W (n): The Hamming window, so the result

will be:

W (n) =0.54-0.46 cos⁡ (2πn/ (N-1)) (3)

0≤n≤N-1

transform:

Step 6: Application of the iDCT (Inverse

Discrete Cosine Transform)

2.2 LPC (Linear Predictive Coding)

LPC analysis is based on the speech

production model mentioned in the figure

below [7]. Starting from the hypothesis

modeling the speech by a linear process, then

It is a linear prediction at an instant n of the p

previous samples. However, the non-linearity

vocal tract and removes the others that do not

parameterization (matrix of the relevant

parameters), or still acoustic vectors extracted

starting from corpus of sounds of the TIMIT

database and from other people with vocal

pathologies, these vectors are injected at the

input of the SVM classifier, the first step is

learning ", and the second step is the test, this

is why the validation base is divided into two

sub-bases one for learning (3/4) and one for

testing (1/4). After a certain number of

-The LPC analysis represents the speech signal

Fig.5. LPC analysis of a few samples of a

Then we present the broad and narrow band

spectrogram

- The broadband spectrogram which is

obtained with a window of short duration (3

ms in our project), it makes it possible to

follow the evolution of the formants, the

voiced periods appear there in the form of dark

bands which are vertical.

- The narrow band spectrogram: it is obtained

Fig.7. Formants display

analysis

Fig.8. PLP and PLP-RASTA analysis

In the following we will present the

parameterization matrices of the different

techniques and establish a comparison

the columns contain the parameter types and

the rows contain the values:

samples of the signal

contains the short-term energy of the

signal.

coefficients.

12 cepstral coefficients plus the pitch

This part groups together the 3 most used

parameterization methods in all that is voice

recognition. Each represented by a matrix.

The dimension of these matrices is 13 columns

and 214 rows. That is 214 frames or vectors

and each with 13 coefficients.

PLP coefficients without the RASTA

filtering.

Normals

It is a question of using an SVM classification

in order to make a distinction between speech

signals coming from people who suffer from

normal signals (no pathology).

In our application, two types of classifications

were used:

- A classification into two classes: in which we

used corpus samples from pathological signals

(Nodule and Oedema) and others from normal

samples from normal signals which is for the

third class. To perform a multi-class

algorithm, i.e. "one against all", this algorithm

consists of taking a signal and comparing it

with all the classes.

1. Learning phase

The learning phase consists of creating a basic

model on which the subsequent classification

of signals is based.

extracted from a signal of a healthy individual

(0) and the green color corresponds when it to

the pathological signals (1)

2. Test phase

The test phase consists of recovering the

matrix resulting from learning in order to

predict or generate a decision, arguments of

SVM classifier:

The following table shows the results obtained

after applying several parameterization

methods, it should be noted that the signals

used include male and female voices.

TIMIT

In this table we have not presented the PLP

RASTA method, because the latter classifies

all the signals as being normal, and this leads

us to say that the PLP RASTA eliminates the

References

processing book (ISNB: 2-88074-388-5)”