Speech reversion detection method and system, terminal equipment and storage medium

Abstract

The invention relates to the technical field of speech recognition detection, and discloses a speech reversion detection method and system, terminal equipment and a storage medium, and the method comprises the steps: carrying out the text representation extraction of an inputted speech feature vector, and obtaining the semantic information of the inputted speech feature vector; carrying out non-text representation extraction on the input voice feature vector to obtain physiological information of the input voice feature vector; carrying out vector splicing on the extracted semantic information and physiological information to serve as an input vector of a k-layer full-connection network, and obtaining an inverse recognition result of an input voice feature vector through training. According to the method, when a general written text is used as the input of an antithreatening detection task, emotion, mood, pause and other non-text features implicit in an original audio are utilized, the two kinds of features are converted into high-level representation, and results are combined for decision making, so that better antithreatening recognition accuracy is achieved.

Description

technical field [0001] The invention belongs to the technical field of voice recognition and detection, and relates to a voice irony detection method, system, terminal equipment and storage medium. Background technique [0002] Irony detection (sarcasm detection) is often regarded as a separate NLP task, generally using written text as input, and analyzing the possibility of using irony in certain words or sentences through contextual association. Because the expression of irony is very obscure, it is difficult to make correct judgments only by focusing on text sentences, so the difficulty of this task is very high, and the actual effect of most current solutions is not ideal. [0003] According to the general practice, if you want to detect irony in people's spoken language, you need to perform a speech recognition task first, convert the audio into text, and then perform subsequent irony detection. In real life, when people use irony, emotions, tone, and pauses will be di...

AI Technical Summary

Problems solved by technology

At present, some researchers use the representation obtained by pre-training the BERT model, or use representations such as GloVe to train the irony detection model, and often can only achieve an F1 value of less than 0.7.

Images