Method for constructing acoustic model, speech recognition system and speech recognition method

Abstract

The invention discloses a method for constructing an acoustic model, comprising: pre-training a gated residual DFSMN neural network module; sequentially connecting a plurality of said gated residual DFSMN neural network modules; The control residual DFSMN neural network module is configured with an input layer and an output layer to form the acoustic module. In the embodiment of the present invention, the gated residual DFSMN neural network module is first pre-trained, and then a plurality of sequential connections are formed to form a deep gated residual DFSMN network as an acoustic model. Since the pre-trained gated residual DFSMN neural network module itself has multiple layers of DFSMN, the obtained acoustic model with multiple gated residual DFSMN neural network modules can contain hundreds of DFSMN layers. In actual training, only a relatively small gated residual DFSMN neural network module needs to be trained, and finally a deep acoustic model is obtained by stacking. Therefore, it not only reduces the cost of training and learning, but also ensures the performance of the final acoustic model.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method for constructing an acoustic model, a speech recognition system and a speech recognition method. Background technique [0002] Speech recognition, in layman's terms, is to convert a piece of speech signal into corresponding text information. Specifically, speech recognition is to sample from a continuous sound wave, quantize each sample value; then divide the quantized sampled audio into frames, and extract a feature vector describing the content of the spectrum for each frame; finally, according to the speech signal The features recognize the words represented by the speech. [0003] The entire process of speech recognition mainly includes feature extraction and decoding (acoustic model, dictionary, language model) parts. [0004] Feature extraction: extract the time-varying speech feature sequence from the speech waveform (that is, convert the sound signal...

AI Technical Summary

Problems solved by technology

However, in practice, it is found that this method of only adding jump links between single layers will still cause the problem of gradient disappearance when training a deep DFSMN with a very large number of layers, resulting in the model not being able to be trained robustly, resulting in Increasing the depth of the model cannot improve the recognition effect, and may even lead to a decrease in the recognition accuracy

Images