Speech emotion recognition based on slice convolution

Abstract

The invention discloses a speech emotion recognition method based on time slicing. The method comprises a first step of speech equal length complementing: setting the maximum time length of a sound and performing time complementing of a read-in speech signal by white noise; a second step of speech signal slicing: slicing the speech signal obtained in the first step, that is, performing the time equal length slicing and speech envelope unit slicing, and the like; a third step of speech segment time-frequency analysis by using a time-frequency analysis tool; a fourth step of sequence reconstruction; a fifth step of emotion marking; a sixth step of image feature extraction; a seventh step of emotion training; an eighth step of speech emotion recognition: obtaining an emotion feature set by using the first to sixth steps for the newly input speech signal, inputting the emotion feature set into an emotion classification model obtained in the seventh step, and finally obtaining the emotion type of the current speech. The method adopts the time slice convolution and deep learning for the speech emotion classification, and well improves the recognition efficiency.

Description

technical field [0001] The invention relates to a voice emotion recognition method based on slice convolution, which belongs to the technical field of computers. Background technique [0002] The extraction of speech emotion features has always been a key issue in speech emotion recognition. At present, the acoustic features of speech emotion recognition can be roughly summarized into prosodic features, spectrum-based correlation features and sound quality features. Among them, prosody refers to the changes in pitch, sound length, speed, and severity that override the semantic symbols in speech. It is a structural arrangement for the expression of speech streams. The commonly used prosody features are sometimes long and fundamental ,Energy. Spectrum-based correlation features are considered to reflect the correlation between vocal tract shape changes and movement. Linear spectra used in speech emotion recognition generally include: LPC, OSALPC, LFPC, etc. Cepstrum features...

AI Technical Summary

Problems solved by technology

[0006] The object of the present invention provides a method for speech emotion recognition based on time slicing. The method slices the speech signal through the time window, and the speech signal after the slice is obtained by time-frequency analysis. Reconstruction of the speech sequence relationship before slicing, this method solves the problem that the speech signal has a long time and cannot be analyzed in a unified time-frequency manner, and uses deep learning for emotion classification, which greatly improves the recognition efficiency

Images