Acoustic model training method and device as well as speech recognition method and device

声学模型训练方法、语音识别方法及装置

Abstract

本发明实施例提供一种声学模型训练、语音识别方法及装置。包括:对训练语料对应的子带进行时域扩展得到特征矩阵;将所述特征矩阵通过预设的卷积器,获得第一向量;根据所述第一向量与预设的隐藏层权重矩阵获得N维向量;其中,N为所述声学模型的分类数目;根据所述N维向量对所述卷积器与所述隐藏层权重矩阵进行参数更新,以所述参数更新后的所述卷积器与所述隐藏层权重矩阵作为所述声学模型。改善了声学模型,提高了声学打分的正确率。
Embodiments of the invention provide an acoustic model training method and device as well as a speech recognition method and device. The acoustic model training method comprises the following steps: carrying out time domain extension on a sub-band corresponding to a training corpus to obtain a characteristic matrix; enabling the characteristic matrix to pass through a preset convolver so as to obtain a first vector; obtaining an N-dimensional vector according to the first vector and a preset hidden layer weight matrix, wherein N is a classification number of an acoustic model; carrying out parameter updating on the convolver and the hidden layer weight matrix according to the N-dimensional vector, and taking the convolver and the hidden layer weight matrix after the parameter updating as the acoustic model. According to the acoustic model training method and device as well as the speech recognition method and device, the acoustic model is improved and the correctness of acoustic scoring is improved.

Claims

Description

Topics

Download Full PDF Version (Non-Commercial Use)

Patent Citations (0)

    Publication numberPublication dateAssigneeTitle

NO-Patent Citations (0)

    Title

Cited By (0)

    Publication numberPublication dateAssigneeTitle