论文部分内容阅读
为了克服低数据资源条件下的资源匮乏问题,该文利用无监督的声学模型训练方法来增加训练数据,改善系统性能。在标准的无监督训练框架下,在传统词图后验概率的词置信度基础上,提出了基于句子后验概率的置信度数据筛选准则,所选数据在保证整句话可靠性的同时很好保留了上下文信息,有利于跨词的三音子声学模型建模;还提出了基于音素覆盖率准则的数据筛选方法,在考虑假设标注句子置信可靠度的同时,尽可能选取训练样本中最为稀疏的音素单元,从源头再次克服低数据资源的困难,数据选择效率更高,性能进一步提升。实验表明:基于本文改进的无监督训练方法的词错误率比基线有监督训练方法的降低约相对8%,比传统无监督方法的也有绝对2%的减少,极大程度改善了低数据资源条件下的系统性能。
In order to overcome the shortage of resources under the condition of low data resources, this paper uses unsupervised acoustic model training methods to increase training data and improve system performance. Under the standard unsupervised training framework, based on the word confidence of the posteriori probability of the traditional word graph, the criteria of confidence data screening based on the posterior probability of the sentence is proposed. The selected data is very reliable Good preserving of contextual information facilitates the modeling of transliteration of triphone acoustic models. A method of data screening based on phoneme coverage criteria is also proposed. While considering the confidence of sentence assignment, Sparse phonetic units, again from the source to overcome the difficulties of low data resources, data selection more efficient and further improve performance. Experiments show that the word-error rate based on the improved unsupervised training method is about 8% lower than that of the baseline supervised training method and an absolute 2% reduction from the traditional unsupervised method, which greatly improves the low data resource conditions Under the system performance.