Speech emotion detection method and apparatus, computer device, and storage medium

A speech emotion detection system, comprising a hardware processor, the hardware processor configured to:

• extracting speech features of to-be-detected speech data to form a speech feature matrix corresponding to the to-be-detected speech data;
• input the speech feature matrix to an emotion state probability detection model, the emotion state probability detection model being trained based on a deep neural network (DNN) model;
• generate an emotion state probability matrix that corresponds to the to-be-detected speech data using the DNN model, the emotion state probability matrix comprising a silent probability which is a probability of a pause in the to-be-detected speech data;
• generate, based on the emotion state probability matrix, the speech feature matrix, and an emotion state transition model, an emotion state sequence corresponding to the to-be-detected speech data, the emotion state sequence comprising at least one silent state, wherein a detection of the at least one silent state does not rely on a mechanism that is based on a power of the to-be-detected speech data;
• detect M successive silent frames according to the at least one silent state in the emotion state sequence, M being an integer;
• in a determination that M is greater than a predefined threshold, segment the to-be-detected speech data according to the M successive silent frames to obtain non-silent speech sub-segments;
• and
• determine, based on the emotion state sub-sequences corresponding to the non-silent speech sub-segments, emotion states corresponding to the to-be-detected speech data.

A non-transitory storage medium for storing computer readable instructions, the computer readable instructions, when executed by a processor in a device, causing the processor to:

• extract speech features of to-be-detected speech data to form a speech feature matrix corresponding to the to-be-detected speech data;
• input the speech feature matrix to an emotion state probability detection model, the emotion state probability detection model being trained based on a recurrent neural network (RNN) model;
• based on the speech feature matrix and the RNN model, generate an emotion state probability matrix corresponding to the to-be-detected speech data, the emotion state probability matrix comprising a silent probability which is a probability of a pause in the to-be-detected speech data;
• generate, based on the emotion state probability matrix, the speech feature matrix, and an emotion state transition model, an emotion state sequence corresponding to the to-be-detected speech data, the emotion state sequence comprising at least one silent state, wherein a detection of the at least one silent state does not rely on a mechanism that is based on a power of the to-be-detected speech data;
• detect M successive silent frames according to the at least one silent state in the emotion state sequence, M being an integer;
• in a determination that M is greater than a predefined threshold, segment the to-be-detected speech data according to the M successive silent frames to obtain non-silent speech sub-segments;
• and
• determine, based on the emotion state sub-sequences corresponding to the non-silent speech sub-segments, emotion states corresponding to the to-be-detected speech data.