| Summary: | The paper presents the results obtained in the interactive phonetic transcription for the Romanian vowels [a], [e] and [i]. After a brief introduction, the resource creation process is presented. A subset of video recordings are selected from the site of Atlasul lingvistic audiovizual al Bucovinei (ALAB) [Audiovisual Linguistic Atlas of Bucovina]. The audio parts, extracted and manually annotated using the Praat software, were phonetically transcribed using the ALR_IIT editor. These result in six sets of feature vectors generated using F0-F3 formant values, MFCC (Mel-Frequency Cepstral Coefficients) coefficients and PLP (Perceptual Linear Prediction) coefficients. For the recognition we used two discriminative classification algorithm: k-NN, for k=1, k=3 and k=5, and SVM. For the k-NN algorithm, we used three distances: Euclidian, Manhattan and Canberra. The results obtained for [a], [e] and [i] are then discussed in detail. The last section is dedicated to the conclusions and to the prospects of our research.
|
|---|