Abstract— This paper makes a comparison of two methods in speech processing that aims to obtain cepstral coefficients that represent characteristics of each speaker, in order to obtain a greater difference between each speaker (to build a representative model, in a next step), traditionally feature extraction is performed through a single periodic window function (generally hamming windows), we compare the traditional method with multi-taper, which has been used in verification speakers tasks and language recognition. The results obtained using 40 speakers indicate a reduction of the variance by using multi-taper and better separation between speakers; with this we have more robust coefficients. To feature extractions we use Mel Frequency Cepstral Coefficients.
Index Terms— Feature extraction, mel frequency cepstral coefficients, multi-taper.
Jose Trangol and Abel Herrera are with Department of Signal Processing, Engineering School, Universidad Nacional Autonoma de Mexico, Mexico (e-mail: abelherrerac1@gmail.com).
[PDF]
Cite: Jose Trangol and Abel Herrera, " Traditional Method and Multi-Taper to Feature Extraction Using Mel Frequency Cepstral Coefficients," International Journal of Information and Electronics Engineering vol. 5, no. 1, pp. 27-30, 2015.