Attention mechanism and convolutional neural network-based voice depression recognition method

Abstract

The invention relates to an attention mechanism and convolutional neural network-based voice depression recognition method. The method comprises the following steps: firstly, preprocessing voice data,and segmenting relatively long voice data according to a standard that a segmented fragment can fully include depression-related features; then extracting a Mel spectrum diagram of each segmented fragment, and adjusting the size of a spectrum diagram input into a neural network model to facilitate model training; then finely adjusting a weight value by a pre-trained Alexnet deep convolutional neural network, and extracting a superior voice feature in the Mel spectrum diagram; performing weight adjustment on fragment-level voice features by using an attention mechanism algorithm to obtain sentence-level voice features; finally, performing depression classification on the sentence-level voice features with an SVM (support vector machine) classification model. The invention provides a novelmethod of depression recognition based on a voice in consideration of the extraction of the depression-related voice features.

Description

technical field [0001] The invention relates to the fields of speech processing, machine learning and deep learning, in particular to a speech depression recognition method based on an attention mechanism and a convolutional neural network. Background technique [0002] Depression is one of the most common emotional disorders, often manifested in low mood, negative attitude, self-blame and other negative states. Depression not only causes harm to oneself, but also has a great impact on daily life, social work, and interpersonal relationships. But at this stage, the diagnosis of depression still depends on the subjective judgment of doctors, and some evaluation scales are used as auxiliary means. Therefore, it is difficult to diagnose depression accurately, making it difficult for patients with depression to receive basic treatment. How to let the computer automatically analyze and judge the severity of the speaker's depression through the speech signal, that is, speech depr...

AI Technical Summary

Problems solved by technology

But the problem is that the above feature extraction process extracts the low-level manually extracted speech features, and does not extract the deeper features in the speech signal, so it cannot fully represent the speech data.

However, the problem is that not all speech clips contain features related to depression, such as silent segments, silent segments

At present, in the field of speech depression recognition, there is no method that uses the combination of attention mechanism and convolutional neural network to identify depression

Images