Impaired Speech Evaluation using Mel-Cepstrum Analysis
Ovidiu Grigore, Corina Grigore, Valentin Velican · 2010
The study presents a simple solution for identifying impaired speech pronunciations using the Mel-Cepstrum Coefficients as features. The pronunciation defect studied was rhotacism - wrongly pronounced 'r' - in the case of Romanian language. By comparing the timbre variation over the duration of the analyzed speech an average of 78% correct classification rate has been obtained. phoneme and the solution used to identify the differences in pronunciation. A brief problem formulation is subsequently presented, outlining the need for such an application and defending the reasons that make this application useful. Then, the method of identifying the mispronunciations is detailed, beginning with a highlight of the most important steps that were followed in developing the application. The core section of the article describes the implementation, presenting the mathematical tools used (Mel-Cepstrum) and the feature extraction method we identified. At the end the method was tested on a small batch of subjects and the results are presented in section V of the article.