论文部分内容阅读
在实际应用中,由于不同领域的数据分布是非常不均匀的,这就导致了不同领域的训练数据与测试数据的之间的识别性能急剧下降。同时,在自适应数据较稀疏的情况下,重训练容易出现过拟合的现象。本文针对不同领域相关的自动语音识别任务,通过深度神经网络构建模型参数共享的框架来进行声学模型训练,同时采用自适应方法来训练跨领域的语言模型。实验结果表明:自适应模型有效地减少了声学模型以及语言模型与识别任务之间跨领域的差异,使得系统的识别率得到有效的提升。
In practice, the distribution of data in different fields is very uneven, which results in a sharp drop in the recognition performance between training data and test data in different fields. In the meantime, under the circumstance that the adaptive data is sparse, the phenomenon of over-fitting is prone to occur in the training. In this dissertation, aiming at the tasks of automatic speech recognition in different fields, this paper constructs a framework for model parameter sharing through deep neural networks to train acoustic models and uses adaptive methods to train cross-domain language models. The experimental results show that the adaptive model effectively reduces the cross-field differences between the acoustic model and the language model and the recognition task, so that the recognition rate of the system is effectively improved.