AI-Powered Speech Recognition for Parkinson's Disease Classification

This project presents a speech recognition system designed to classify individuals as either healthy or affected by Parkinson's disease based on their voice characteristics. The project utilizes a dataset of speech recordings and employs a shallow neural network classification technique implemented in MATLAB. The methodology involves loading the dataset, applying different algorithms with varying parameters, and selecting the optimal algorithm (Bayesian Regularization back-propagation) based on performance. Experiments were conducted, and the results indicate a satisfactory classification accuracy, with over 73% of cases correctly identified. The project discusses the limitations of the system, such as its inability to handle multiple classes or directly process vocal samples without pre-processing. The report includes details on the dataset, methodology, experimental results, and MATLAB code, along with plots of training state, error histograms, ROC curves, and confusion matrices. The project highlights the successful application of AI algorithms in identifying Parkinson's disease through speech analysis, while also acknowledging areas for potential improvement and future research.

Speech Recognition Using Shallow
Neural Network Classification
Introduction:
Speech recognition is a subfield of computer linguistics by which methodologies are
developed by which enables the machines to recognize the spoken language by
human by converting into text.
In this particular research a different type of speech recognition application is
performed where the by different attributes of speech a person is identified as a
healthy or diseased. In particular the recognition software will be able to recognize
when a diseased person (for a particular disease) speaking to the system by analysing
its attributes or when a healthy person is speaking. Now, instead of considering
multiple disease detection the software will be implemented for only one disease
which is chosen to be Parkinson’s disease where the patient goes through voice
change with other change in physical attribute change.

Problem definition:
The problem can be defined as building an automatic artificial
intelligence system that can classify a person as health or affected by
Parkinson’s disease when the key attributes of the voice sample is given
as input to the system in appropriate format.

Dataset information
Parkinson Speech Dataset with Multiple Types of Sound Recordings
Data Set.
Available at:
https://archive.ics.uci.edu/ml/datasets/Parkinson+Speech+Dataset+wit
h++Multiple+Types+of+Sound+Recordings#
Number of instances: 1040
Number of variables: 26
Variable type: Integer, numeric
Different test and train set

Methodology:
Entire dataset containing training and testing data are loaded in a
programming software MATLAB.
MATLAB is chosen as this has many inbuilt libraries and functions for
neural networks and artificial intelligence which are needed for training
and testing with the chosen Parkinson’s data.
Different algorithms are applied with different parameters such as
number of hidden layer and initial weights of bias vectors. The
algorithm with best performance with optimal parameter is chosen for
classification.

Experiments and discussion:
Many testing of the initial neural network is performed in sequence to
improve the performance on test case.
Chosen performance function: mean square error
Optimum performance with: Bayesian Regularization back-propagation
method (‘trainbr’)
Hidden layer size = 50

Bayesian regularization
algorithm:
Bayesian regularized ANN are more efficient than the normal back-
propagation nets and thus the need of lengthy cross-validation can be
reduced by this method.
Main advantage of algorithm: Robust and validation not required which
save O(N^2) time complexity.
Overtraining in this method is very difficult as an objective Bayesian
criterion is required to stop the training. The over fitting to the data is
also quite unlikely as ‘trainbr’ calculates and then trains on a wide
variety of network parameters and weights of each node that
eliminates the nodes which are not relevant.

Bayesian regularization
algorithm:
The effective number of nodes or neurons in BR method are generally
very small compared to the total number of weights in a complete
back-propagation neural network.
When employed with ARD method then it helps to filter out the
irrelevant or highly correlated variables.
Only mean squared error or the sum of square error can be applied in
BR as performance function.
The BR algorithm is basically an extended version of the Levenberg-
Marquardt algorithm where the back-propagation is used for
calculating the Jacobian of the performance corresponding to bias
variables X and the weights.

Mathematical formulation of
algorithm
• F = objective function
• M = weights of neural network
• e = error in modelling
• w = net weight
• D = Input-output combination of dataset
• H = Hessian of the objective function calculated using Jacobian
•

Mathematical formulation of
algorithm
Objective: Min, F =
Equivalent to, maximize likelihood function
(By Bayes theorem)
Similarly,
Gaussian function

Mathematical formulation of
algorithm
likelihood function
=
Thus after calculating all the probabilities the can easily be calculated.

Training of algorithm stopping
conditions:
• maximum number of specified epochs are reached
• maximum time specified is reached in seconds
• required performance is reached
• performance gradient is under minimum gradient as specified
• Marquardt adjustment parameter μ becomes more than maximum
value of adjustment parameter μmax.

Default BR parameters in MATLAB
• Max epoch number = 1000
• Performance objective value (mse) = 0
• Marquardt adjustment parameter (μ)= 0.005
• Decrement factor of μ = 0.1
• Increment factor of μ= 10
• Maximum value of μ = 10^10
• Maximum number of validation fails = ∞
• Minimum performance gradient = 10^-7
• Maximum training time = ∞