Emotional Recognition Using a Compensation Transformation in Speech Signal

Computational Linguistics and Chinese Language Processing Vol. 12, No. 1, March 2007, pp. 79-90 79  The Association for Computational Linguistics and Chinese Language Processing [Received August 11, 2006; Revised September 30, 2006; Accepted October 22, 2006] Emotional Recognition Using a Compensation Transformation in Speech Signal Cairong Zou∗ , Yan Zhao+, Li Zhao +, Wenming Zhen+, and Yongqiang Bao+ Abstract An effective method based on GMM is proposed in this paper for speech emotional recognition; a compensation transformation is introduced in the recognition stage to reduce the influence of variations in speech characteristics and noise. The extraction of emotional features includes the globe feature, time series structure feature, LPCC, MFCC and PLP. Five human emotions (happiness, angry, surprise, sadness and neutral) are investigated. The result shows that it can increase the recognition ratio more than normal GMM; the method in this paper is effective and robust. Key words: Speech Emotional Recognition (SER), GMM, Emotion Recognition, Compensation Transformation 1. Introduction One of the natural goals for research on speech signals is recognizing emotions of humans [Chen 1987; Oppenheim 1976; Cowie 2001]; it has gained growing amounts of interest over the last 20 years. A study conducted by Shirasawa et al. showed that SER could be made by ICA and attain an 87% average recognition ratio [Shirasawa 1997; Shirasawa 1999] Many studies have been conducted to investigate neural networks for SER. Chang-Hyun Park tried to recognize sequentially inputted data using DRNN in 2003[Park et al. 2003], Muhammad, W. B. obtained about 79% recognition rate using GRNN [Bhatti et al. 2004]. Aishah Abdul Razak achieved an average recognition rate of 62.35% using combination MLP [Razak et al. 2005]. Fuzzy rules are also introduced into SER such that an 84% rate has been achieved in recognizing anger and sadness [Austermann et al. 2005]. A number of studies in SER have ∗ Foshan University, Foshan, 528000, Guangdong, China + Research Center of Learning Science, Southeast University, Nanjing, 210096, China E-mail: zhaoli@seu.edu.cn 80 Cairong Zou et al. also been done with the development of GMM/HMM [Rabiner 1989; Jiang et al. 2004; Lin et al. 2005]. However, in SER, the variations in speech characteristics, noise and individual differences always influence the recognition results. In addition, the methods above have always handled such problems in the preprocessing stage and have not been able to eliminate the influence effectively. Therefore, a valid solution has still not been proposed. In this paper a compensation transformation is introduced into an algorithm for GMM which operates in the recognition module. The experiments with five emotions (happiness, angry, neutral, surprise and sadness) show that the method in this paper is effective in emotional recognition. 2. Descriptions of Emotion and Selection of Emotion Speech Materials Usually, emotions are classified into two main categories: basic emotions and derived emotions. Basic emotions, generally, can be found in all mammals. Derived emotions mean derivations from basic emotions. One viewpoint is that the basic emotions are composed by the basic mood. Due to different research backgrounds, different researchers …

Content truncated to 500 words