DOI 10.17586/0021-3454-2015-58-6-443-460
UDC 004.934.1
DEVELOPMENT OF EMOTION-TOLERANT INFORMATIVE INDICATORS FOR SPEECH RECOGNITION PROBLEM
Speech Technology Ltd., Minsk; Junior Scientist
Read the full article
Abstract. A method of the speech signal parameterization providing emotion-tolerant and speaker-invariant feature vector is proposed. The method makes use of the cepstral coefficients defined on an ExpoLog frequency scale on the base of on a linear-prediction power spectrum. The described feature vector is applied for emotional speech recognition based on hidden Markov models. Experimental results demonstrate that the use of the proposed method improves emotional speech recognition efficiency by 5,9 %.
Keywords: emotional speech recognition, feature vector, linear prediction coefficients, cepstral coefficients, hidden Markov models
