论文部分内容阅读
为了提高噪声环境下语音识别系统的鲁棒性,本文提出了一种基于迁移学习的声学建模方法。该方法用干净语音的声学模型(老师模型)指导带噪语音的声学模型(学生模型)进行训练。学生模型在训练过程中,尽量使其逼近老师模型的后验概率分布。学生模型和老师模型间的后验概率分布差异通过相对熵(KL Divergence)加以最小化。据CHiME-2数据集上的实验结果表明,本文所提方法的平均词错误率(WER)与基线相比绝对下降了7.29%,与CHiME-2竞赛的第一名相比绝对下降了3.92%。
In order to improve the robustness of speech recognition system under noisy environment, this paper presents a method of acoustic modeling based on migration learning. This method uses a clean acoustic model (teacher model) to guide the noisy speech acoustic model (student model) to train. In the training process, the student model try to approximate the posterior probability distribution of the teacher model. The posterior probability distribution differences between the student model and the teacher model are minimized by the relative entropy (KL Divergence). According to the experimental results on the CHiME-2 dataset, the average word-error rate (WER) of the proposed method dropped by an absolute 7.29% from the baseline and declined by an absolute 3.92% compared with the first place in the CHiME-2 competition .