Speech enhancement method and system fusing signal-to-noise ratio and intelligibility dual targets

Abstract

The invention belongs to the technical field of speech enhancement signal processing, and particularly relates to a speech enhancement method integrating a signal-to-noise ratio and intelligibility dual targets. The method comprises the following steps: converting an original speech signal into an original time-frequency domain feature; inputting the original time-frequency domain feature into a pre-established first neural network model to obtain a first effective feature with a signal-to-noise ratio; inputting the original time-frequency domain features into a pre-established second neural network model to obtain a second effective feature with intelligibility; processing the first effective feature and the second effective feature to obtain a weight matrix, selecting an element with high correlation with the first effective feature from the second effective feature column by column from the weight matrix according to a preset correlation weight threshold, extracting a correlation weight threshold of the element, replacing a threshold value at a corresponding position in the first effective feature with the correlation weight threshold, taking the replaced first effective featureas the time-frequency domain feature after speech enhancement, and converting the time-frequency domain feature after speech enhancement into an enhanced speech signal.

Description

technical field [0001] The invention belongs to the technical field of speech enhancement signal processing, and in particular relates to a speech enhancement method and system that integrates dual objectives of signal-to-noise ratio and intelligibility. Background technique [0002] When the speech signal is interfered by noise, the signal quality and intelligibility will decrease, thereby affecting the user experience of speech recognition and speech perception processing based on the speech signal. At present, the commonly used speech enhancement methods rely on estimating the mask of the speech signal, and then separate the spectral components of the speech signal from the noise coverage. This speech enhancement method is usually based on the minimum mean square error criterion, estimates a mask, classifies the speech signal components with noise in the time-frequency domain, distinguishes the components covered by noise, and retains the components with strong speech sig...

AI Technical Summary

Problems solved by technology

For the various optimization objectives faced by existing speech enhancement methods, a single training criterion cannot comprehensively cover the errors obtained from all optimization objective angles, and usually only achieves the balance between suppressing noise residue and improving auditory quality and intelligibility. Balanced, the enhancement results are not optimal in terms of suppressing noise and improving speech intelligibility

Images