Elżbieta Smołka

Automatic Detection of Disorders in a Continuous Speech with the Hidden Markov Models Approach

Hidden Markov Models are widely used for recognition of any patterns appearing in an input signal. In the work HMM’s were used to recognize two kind of speech disorders in an acoustic signal: prolongation of fricative phonemes and blockades with repetition of stop phonemes.

Effect of Acoustical, Visual and Tactile Echo on Speech Fluency of Stutterers

The study presents the comparison of the effects of echo transmitted via single and combined channels (auditory, visual and tactile) on the speech of stutterers. The dependence of stuttering intensity and speech velocity upon echo delay time was determined. For all transmission channels the stuttering intensities and the speech velocities decreased with the increase in the delay time of the echo. The results were analyzed statistically by means of the ANOVA method. It was proven that the corrective effects of visual echo and tactile echo were comparable. Echo transmitted via the auditory channel was more effective than when transmitted via the visual or tactile channels. The greatest efficiency could be observed by transmitting echo via three connected channels: auditory, visual and tactile. The results obtained show that in stuttering therapy it is justified to use echo transmitted via three connected channels (auditory, visual, tactile).

Effect of Acoustical, Visual and Tactile Reverberation on Speech Fluency of Stutterers

The study presents the comparison of the effects of reverberation transmitted via single and combined channels (auditory, visual and tactile) on the speech of stutterers. The dependence of stuttering intensity and speech velocity upon reverberation time was determined. For all transmission channels the stuttering intensities and the speech velocities decreased with the increase in reverberation time. The results were analyzed statistically by means of the ANOVA method. It was proven that the corrective effects of visual reverberation and tactile reverberation were comparable. Reverberation transmitted via the auditory channel was more effective than when transmitted via the visual or tactile channels. Connecting the visual and tactile channels with the auditory channel has no influence on the effectiveness of reverberation.

Speech nonfluency detection using Kohonen networks

This work covers the problem of application of neural networks to recognition and categorization of non-fluent and fluent utterance records. Fifty-five 4-s speech samples where the blockade on plosives (p, b, t, d, k and g) occurred and 55 recordings of speech of fluent speakers containing the same fragments were applied. Two Kohonen networks were used. The purpose of the first network was to reduce the dimension of the vector describing the input signals. A result of the analysis was the output matrix consisting of the neurons winning in a particular time frame. This matrix was taken as an input for the next self-organizing map network. Various types of Kohonen networks were examined with respect to their ability to classify utterances correctly into two, non-fluent and fluent, groups. Good examination results were accomplished and classification correctness exceeded 76%.

Artificial Neural Networks in the Disabled Speech Analysis

Presented work is a continuation of conducted research concerning automatic detection of disfluency in the stuttered speech. So far, the experiments covered analysis of disorders consisted in syllable repetitions and blockades before words starting with stop consonants. Introduced work gives description of an artificial neural networks application to the recognition and clustering of prolongations, which are one of the most common disfluency that appears among stuttering people.The main aim of the research was to answer a question whether it is possible to create a model built with artificial neural networks that is able to recognize and classify disabled speech. The experiment proceeded in two phases. In the first stage, Kohonen network was applied. During the second phase, two various networks were used and next evaluated with respect to their ability to classify utterances correctly into two, non-fluent and fluent, groups.

Prolongation detection with application of fuzzy logic

Due to great value of the time constant of the integrator circuit, a hardware defined PWM (Pulse Width Modulation) signal makes it possible to build a Digital to Analog Converter (DAC) characterized by a relatively long response time. An attempt at creating the software defined PWM signal leads to increased response time of the DAC converter with a coefficient several hundred times longer than its hardware equivalent. Reorganization of the PWM signal allows for its software synthesis, in that the response time of the DAC converter is only several times larger than its classical equivalent.In the paper we present short profile of different compositions of many-layered applications. We compare speed and stability of the different sets of applications. We describe a configuration which can be used to create a server of scientific information. The application consists of MySQL database management system, http server (Apache) and PHP language. All operate in the Linux environment.

The Prolongation-Type Speech Non-fluency Detection Based on the Linear Prediction Coefficients and the Neural Networks

The goal of the paper is presenting a speech prolongation detection method based on the linear predicion coefficients obtained by the Levinson-Durbin method. The application “Dabar”, which was made for this aim, has an ability of setting the coefficients computed by the implemented methods as an input of the Kohonen networks with different size of the output layer. Three different types of the neural networks were used to classify fluency of the utterances: RBF networks, linear networks and Multi-Layer Perceptrons. The Kohonen network (SOM) was used to reduce the LP coefficients representation to the winning neurons vector. After that the vector was splitted into subvectors whom represents 400ms utterances. These utterances were fragments of the Polish speech without the silence. The research was based on 202 fluent utterances and 140 with the prolongations on Polish phonems. The classifying success reached 75% of certainty.

Articulation Rate Recognition by Using Artificial Neural Networks

This works concerns the problem of the application of artificial neural networks in the modelling of the hearing process. The aim of the research was to answer the question whether artificial neural networks are able to evaluate speech rate. Speech samples, first recorded during reading of a story with normal and next with slow articulation rate were used as research material. The experiment proceeded in two phases. In the first stage Kohonen network was used. The purpose of that network was to reduce the dimensions of the vector describing the input signals and to obtain the amplitude-time relationship. As a result of the analysis, an output matrix consisting of the neurons winning in a particular time frame was received. The matrix was taken as input for the following networks in the second phase of the experiment. Various types of artificial neural networks were examined with respect to their ability to classify correctly utterances with different speech rates into two groups. Good examination results were accomplished and classification correctness exceeded 88%.

Automatic Disordered Syllables Repetition Recognition in Continuous Speech Using CWT and Correlation

Automatic disorder recognition in speech can be very helpful for the therapist while monitoring therapy progress of the patients with disordered speech. In this article the syllables repetition are described. The signal was analyzed using Continuous Wavelet Transform with bark scales, the result was divided into vectors (using windowing) and then a correlation algorithm was used on this data. Quite large search analysis was performed during which, recognition above 80% was achieved. All the analysis was performed and the results were obtained using the authors program – WaveBlaster. It is very important that the recognition ratio above 80% was obtained by a fully automatic algorithm (without a teacher) from the continuous speech. The presented problem is part of our research aimed at creating an automatic disorders recognition system.

Visual feedback in stuttering therapy

The aim of this paper is to present the results concerning the influence of visual echo and reverberation on the speech process of stutterers. Visual stimuli along with the influence of acoustic and visual-acoustic stimuli have been compared. Following this the methods of implementing visual feedback with the aid of electroluminescent diodes directed by speech signals have been presented. The concept of a computerized visual echo based on the acoustic recognition of Polish syllabic vowels has been also presented. All the research nd trials carried out at our center, aside from cognitive aims, generally aim at the development of new speech correctors to be utilized in stuttering therapy.