Training and identification method and system of bidirectional neural network model

Abstract

The invention discloses the training and identification method and system of a bidirectional neural network model used for processing a noisy voice. The method comprises the followings steps of acquiring simulation noisy data and real noisy data; calculating the time-frequency hidden value label of the simulation noisy data, setting the label of the simulation noisy data to be a training target ina bidirectional neural network, and inputting the simulation noisy data after preset processing into the bidirectional neural network so as to carry out training; using a clustering mode to estimatethe soft time-frequency hidden value label of the real noisy data, setting the soft label of the real noisy data in the bidirectional neural network to be the training target, and inputting the real noisy data after preset processing into the bidirectional neural network so as to carry out training; and outputting the neural network parameters of the trained bidirectional neural network. In the invention, through introducing the real and non-simulated training data, the neural network model is trained. On one hand, a training data size is increased; on the other hand, the mismatching of the simulation data and the real data is reduced.

Description

technical field [0001] The invention belongs to the field of speech signal processing, and in particular relates to a training and recognition method, system and electronic equipment of a bidirectional neural network model for processing noisy speech. Background technique [0002] In recent years, significant progress has been made in Automatic Speech Recognition (ASR) due to the introduction of deep neural networks into acoustic modeling. ASR systems based on deep neural networks still perform poorly in many real-world far-field microphone scenarios. The main reason for poor performance is background interference, such as additive noise, channel distortion and reverberation, which reduces the signal-to-noise ratio and degrades the performance of ASR. Beamforming has proven to be a useful front-end approach to improve system performance under these conditions. While traditional beamforming methods usually rely on inaccurate prior knowledge, such as array geometry or plane ...

AI Technical Summary

Problems solved by technology

[0004] In the process of realizing the present invention, the inventors found that although the neural network-based method superficially surpasses the cluster-based method, it is easy to cause a mismatch between training and testing conditions

This mismatch arises because NN-based hidden value estimators can only be trained on parallel simulated data implemented with real data

Images