Speech enhancement method based on hybrid masking learning target

Abstract

A speech enhancement method based on a hybrid masking learning target comprises the steps of performing traditional feature extraction of speech signals, dividing the acquired speech signals into a training set and a test set, and respectively extracting traditional features of the speech signals of the training set and the test set, respectively extracting amplitude spectrum features of STFT domains of the speech signals of the training set and the test set, constructing a deep stacking residual network, constructing a learning target, training the deep stacking residual network by using theextracted traditional features of the training set, the extracted amplitude spectrum features of the STFT domain and the learning target, and inputting the extracted traditional features of the test set and the amplitude spectrum features of the STFT domain into the trained deep stacking residual network to obtain a predicted learning target, performing ISTFT on the predicted learning target to obtain an enhanced speech signal, and calculating a PESQ value of the speech signal. Noise information is not reserved in the speech-dominated time frequency unit, so that the calculation amount is reduced, and neural network learning is easy to train to improve the intelligibility and quality of speech.

Description

technical field [0001] The present invention relates to a hybrid masking learning objective. In particular, it concerns a method for speech enhancement based on hybrid masking learning objectives. Background technique [0002] At present, there are many speech enhancement methods based on deep learning, and the key technologies mainly involve three aspects: which feature to extract, which model to use, and which target to learn. Like features, the study of learning objectives is also very valuable. Under the premise of the same training data characteristics and learning model, better learning objectives can make model training better. [0003] In a speech enhancement system using a supervised neural network, the acquisition of learning targets is generally calculated based on background noise and pure speech. Effective learning targets have an important impact on the learning ability of the speech enhancement model and the generalization of the system. [0004] Currently u...

AI Technical Summary

Problems solved by technology

Commonly used time-frequency masking targets include: Ideal Binary Mask (IBM), Ideal Ratio Mask (IRM), Target Binary Mask (TBM), etc.; among them, the most commonly used The learning objectives are ideal binary masking and ideal floating value masking, but these two learning objectives have their own shortcomings such as inaccurate prediction and poor generalization.

[0005] When the learning target is IRM, the model only needs to classify (0 or 1) whether each time-frequency unit is dominated by noise or target voice, which will cause noise information to be retained in the time-frequency unit dominated by the target voice, and these noise signals will be Seriously affect the intelligibility and quality of speech; when the learning target is IRM, the model needs to predict the coefficients in each time-frequency unit. In the time-frequency unit dominated by noise, the extracted features cannot well represent this The characteristics of the target speech in the time-frequency unit, but for the model, it is difficult to accurately predict the coefficient of the time-frequency unit

Images