Speech enhancement method

Abstract

The invention discloses a speech enhancement method. The speech enhancement method comprises the following steps of extracting acoustic features of each speech frame; training a progressive dual-output neural network model by using samples of the clean speech and the noise speech, estimating ideal soft masking of each speech frame by using the trained progressive dual-output neural network model,and performing enhancement processing of the acoustic features; if the speech enhancement method is applied to the human ears, reconstructing a waveform by using the enhanced acoustic features to obtain a waveform that can be subjected to subjective audiometry; and if the speech enhancement method is applied to a speech recognition system, then applying the estimated ideal soft masking to acousticfeatures of the input speech to obtain the masked acoustic features, and then reconstructing the waveform to obtain the enhanced speech. The above scheme of the invention can meet the noise reductionrequirements of the human ears and improve the recognition accuracy of the noisy speech.

Description

technical field [0001] The invention relates to the technical field of speech processing, in particular to a speech enhancement method. Background technique [0002] Speech recognition is to let the machine understand what people say, that is, to convert the vocabulary content in human speech into input that the computer can recognize. In the past 20 years, especially the introduction of deep learning in recent years, speech recognition technology has achieved remarkable results, and it has begun to move from the laboratory to the market. At present, speech input, speech retrieval, and speech translation based on speech recognition technology have been widely used. As we all know, if we do not take any measures in the noisy environment, the performance of automatic speech recognition will drop significantly, the main reason is the difference between the noisy speech distribution and the acoustic model distribution. In order to improve the recognition accuracy in noisy envi...

AI Technical Summary

Problems solved by technology

The main problem with NMF-based speech enhancement methods is that the models for noisy and clean speech are trained separately and independence between noisy and clean speech is assumed

In addition, the training of the neural network is learned off-line. Like a human, it can remember some noise patterns, so it can suppress some non-stationary noise well; however, if the training data does not match the test data, for example, the noise type is different. , the speakers have a large difference, and the system performance will drop significantly

Images