Acoustic modeling method and device for multi-channel speech recognition based on spatial feature compensation

Abstract

The present invention relates to a multi-channel speech recognition acoustic modeling method and device based on spatial feature compensation. The proposed model is based on the traditional hybrid acoustic modeling framework, that is, the neural network acoustic model predicts the state posterior probability of the hidden Markov model, The method comprises: extracting the acoustic features and spatial information features in the microphone array of each single-channel recorded voice signal in the microphone array; inputting the acoustic features and the spatial information features into the neural network acoustic model training; the neural network The network acoustic model outputs the predicted posterior probability of the acoustic state, and the acoustic model optimization criterion is used to iteratively update the neural network parameters to generate a multi-channel speech recognition acoustic model based on spatial feature compensation. This method avoids the suboptimal solution caused by the separate optimization of the front and rear ends in the traditional method; it enables the neural network acoustic model to effectively use the spatial information provided by the microphone array, and improves the acoustic modeling ability of multi-channel speech signals.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to an acoustic modeling method and device for multi-channel speech recognition based on spatial feature compensation. Background technique [0002] In recent years, acoustic modeling methods based on Deep Neural Network (DNN) have achieved breakthrough results in the field of speech recognition. The introduction of complex neural networks such as Long Short-Term Memory (LSTM) has further improved the ability of acoustic modeling. However, far-field speech recognition tasks are still challenging due to background noise, reverberation, and human voice interference. [0003] Data recorded by multiple microphones can provide additional spatial information compared to single-microphone acquisition of speech signals. Therefore, a microphone array is usually used to improve the recognition accuracy of far-field speech signals. Traditional multi-channel speech recognition systems genera...

AI Technical Summary

Problems solved by technology

Although studies have shown that DNN can directly use multi-channel output in parallel as network input to model the acoustic state posterior probability, this method still does not make full use of the spatial information introduced by the microphone array, that is, the speaker's position information

Images