摘要 |
本发明实施方式提出一种声学模型训练方法和装置。方法包括:建立深层神经网路模型初始模型;将语音训练资料划分为N个不相交的资料子集合,针对每个资料子集合利用随机梯度下降演算法更新深层神经网路模型初始模型,得到N个深层神经网路模型子模型,其中N为至少为2的自然数;融合N个深层神经网路模型子模型以得到深层神经网路模型中间模型,并当该深层神经网路模型中间模型符合预先设定的收敛条件时,判定该深层神经网路模型中间模型为训练后声学模型。本发明实施方式提高了声学模型的训练效率,并且不降低语音识别的性能。; dividing speech training data into N disjoint data subsets, and updating the initial model of deep neural network model for each data subset by using stochastic gradient descent algorithm to obtain N sub-model of deep neural network model, where N is a natural number of at least 2; fusing the N sub-model of deep neural network model to get the intermediate model of deep neural network model, and when the intermediate model of deep neural network model in line with a preset convergence condition, determining the intermediate model deep neural network model as a acoustic model after training. Embodiment of the present invention improves the efficiency of the acoustic model training, without reducing the performance of speech recognition. |