Decision Tree Versus Linear Support Vector Machine Classifier in the Screening of Medial Speech Sounds: A Quest for a Sound Rationale.

This paper describes the latest development in the classification stage of our Speech Sound Disorder (SSD) Screening algorithm and presents the results achieved by using two classifier models: the Classification and Regression Tree (CART)-based model versus the Single Decision Hyperplane-based Linear Support Vector Machine (SVM) model. For every single speech sound in medial position, 10 features extracted from the audio samples along with an 11th feature representing the validation of the (mis)pronunciation by the Speech Language Pathologist (SLP) were fed into the 2 classifiers to compare and discuss their performance. The accuracy achieved by the two classifiers on a data test size of 30% of the analyzed samples was 98.2% for the Linear SVM classifier, and 100% for the Decision Tree classifier, which are optimal results that encourage our quest for a sound rationale.

Support-Vector Machine-Based Classifier of Cross-Correlated Phoneme Segments for Speech Sound Disorder Screening.

This paper presents a Support-Vector Machine (SVM) based method of classification of cross-correlated phoneme segments as part of the development of an automated Speech Sound Disorder (SSD) Screening tool. The pre-processing stage of the algorithm uses cross-correlation to segment the target phoneme and extracts data from the new homogeneously trimmed audio samples. Such data is then fed into the SVM-based classification script which currently achieves an accuracy of 97.5% on a dataset of 132 rows. Given the global context of an increasing trend in the incidence of Speech Sound Disorders (SSDs) amongst early-school aged children (5-6 years old), the constraints imposed by the new Corona virus pandemic, and the (consequent) shortage of professionally trained specialists, an automated screening tool would be of much assistance to Speech-Language Pathologists (SLPs).

Word-Final Phoneme Segmentation Using Cross-Correlation.

The goal of this paper is to present a word-final target phoneme automated segmentation method based on cross-correlation coefficients computed between a reference sound wave and a sample sound wave. Most existing Speech Sound Disorder (SSD) Screening solutions require human intervention to a greater or lesser extent and use segmentation methods based on hard-coded time frames. Moreover, existing solutions extract features from the frequency domain, which entails large amounts of computational power to the detriment of real-time feedback. The pre-processing algorithm proposed in this paper, implemented in a Python version 3.7 script, automatically generates 2 new .wav files corresponding to the phonemes found in word-final position in the initial sound waves. The newly-generated .wav files are meant to be used as valid and homogeneous input in a subsequent classification stage aimed at rigorously discriminating mispronunciations of the target phoneme and assist Speech-Language Pathologists (SLPs) with the SSD screening.

Cross-Correlation Based Automated Segmentation of Audio Samples.

This paper presents an audio file segmentation method in an attempt to mitigate the issue of variable durations of the same utterance by different individuals, e.g.: Speech-Language Pathologist (SLP) and dyslalic subjects. The Method section describes the manner of determination of the maximum cross-correlation value between the 2 audio files and the subsequent automated segmentation thereof in order to extract 2 valid pronunciation samples of the target consonant. The method is aimed at pre-processing audio files and supplying homogeneously-trimmed audio samples to a computerized SSD Screening system. The results obtained on a batch of 30 pronunciations are presented and briefly discussed in the third section while the last section is reserved for conclusions and perspectives.

Optimization of Entropy-Based Automated Dyslalia Screening Algorithm.

This paper presents the current state of progress of a project aimed at achieving an automated information entropy-based discrimination of phoneme mispronunciations in utterances of early school-age children. The introductory part briefly describes the dyslalia symptomology and the incidence of dyslalic disorders. This section also reviews the current challenges posed by the main research objective in other similar projects sharing the same objective and summarizes the current results thereof. The Material and Method section presents the conditions, the technology and the feature-extraction technique used in the experiment. The same section also describes the computation of the information entropy values of each analyzed speech sample. The highest match rate of 93.33% was achieved in the classification of words containing the phoneme /r/ in the initial position. A synthesis of the achieved results is provided in the Results section based on which conclusions are drawn and exposed in the Discussion and Conclusions section.

Entropy-based Dyslalia Screening.

This paper's objective is to present a proposed solution of Computer-based Speech Therapy System (CBST) for dyslalia screening. The problem of Speech Sound Disorders (SSD) is enunciated, and a brief presentation of several general CBST solutions is made. An Entropy-based method is proposed and the current state of advancement in the development and experimental validation of this solution is presented and discussed. Conclusions related to future improvements of the method are drawn based on the consequences identified in the final section.

A Speech Sound Disorder Screening System Database Structure.

This paper makes a brief review of several database structures of Computer-Based Speech Therapy (CBST) systems and solutions and describes the screening method, an experimental study conducted to validate the screening algorithm and a database structure for the Information Entropy-Based Sound Speech Disorder (SSD) Screening System aimed at by our research project. The final part briefly presents the essential design criteria and further development.

A Computer-Based Speech Sound Disorder Screening System Architecture.

This paper reviews several architectures of Computer-Based Speech Therapy (CBST) systems and solutions and describes an architecture for an Entropy-Based Sound Speech Disorder (SSD) Screening System aimed at by our research project. The proposed architecture and data flow scenario aim to provide a fully-automated Entropy-based SSD Screening System, to be connected with CBSTs and to be used as a research infrastructure for further refinement of the objectives of our research project.