Speech-emotion recognition method based on improved quadratic discriminant

Abstract

The invention discloses a speech-emotion recognition method based on improved quadratic discriminant. By utilizing the method, false recognition rate caused by probability distribution diversity of emotional characteristic parameter statistics can be effectively reduced. The method comprises the steps of performing exponential transformation on characteristic parameters to enable parameter distribution after transformation to be near normalized, estimating probability distribution functions of original characteristics on the premise of subjecting transformed parameters to normal distribution, and using a logarithmic form to obtain improved quadratic discriminant. Compared with the prior other characteristic normalization transformation, the exponential transformation adopted by the method can more effectively normalize the characteristic parameters, and recognition rate can be effectively improved by adopting the improved quadratic discriminant.

Description

technical field [0001] The invention relates to a speech recognition method, in particular to a speech emotion recognition system and method. Background technique [0002] The automatic speech emotion recognition technology mainly includes two problems: one is to use the characteristics in the speech signal as emotion recognition, that is, the problem of emotional feature extraction, and the other is how to classify specific speech data, that is, the problem of pattern recognition . [0003] The emotional features commonly used in speech emotion recognition are mainly prosody parameters and sound quality parameters. The former includes duration, speech rate, energy, pitch frequency and its derivative parameters. The sound quality parameters are mainly formants, harmonic-to-noise ratio and their derivative parameters, etc. . However, due to the individual differences between people (the variability of the vocal tract, the characteristics of the vocal tract, the pitch of the...

AI Technical Summary

Problems solved by technology

[0004] Before the present invention, among the various existing identification methods, although the neural network method has a high degree of nonlinearity and strong classification ability, the required learning time increases rapidly with the increase of the network, and the local minimum The problem is also a shortcoming; the hidden Markov method (HMM) takes a long time to establish and train, and it needs to solve the problem of high computational complexity when it is applied in practice

Although the quadratic discriminant algorithm is simple and has a small amount of calculation, it must be based on the premise that the feature vector obeys the normal distribution, which greatly affects the recognition rate. So far, the normalization of the feature vector, such as root or Box-Cox The transformation has a better effect on the parameters of the near Г distribution, and the parameter distribution mentioned above is diverse and non-normal. How to find a more effective normal transformation for its probability function is to use the secondary discriminant for identification. issues that must be considered

Images