论文部分内容阅读
众所周知中文普通话被众多的地区口音强烈地影响着,然而带不同口音的普通话语音数据却十分缺乏。因此,普通话语音识别的一个重要目标是恰当地模拟口音带来的声学变化。文章给出了隐式和显式地使用口音信息的一系列基于深度神经网络的声学模型技术的研究。与此同时,包括混合条件训练,多口音决策树状态绑定,深度神经网络级联和多级自适应网络级联隐马尔可夫模型建模等的多口音建模方法在本文中被组合和比较。一个能显式地利用口音信息的改进多级自适应网络级联隐马尔可夫模型系统被提出,并应用于一个由四个地区口音组成的、数据缺乏的带口音普通话语音识别任务中。在经过序列区分性训练和自适应后,通过绝对上0.8%到1.5%(相对上6%到9%)的字错误率下降,该系统显著地优于基线的口音独立深度神经网络级联系统。
As we all know, Mandarin Chinese is strongly influenced by many regional accents, but Mandarin pronunciation data with different accents is lacking. Therefore, an important goal of Mandarin speech recognition is to properly simulate the acoustical changes brought about by the accent. The article gives a series of research on acoustics model based on deep neural network which implicitly and explicitly uses accent information. At the same time, multi-accent modeling methods including mixed condition training, multi-accent decision tree state binding, deep neural network cascade and cascade Hidden Markov Model modeling of multi-level adaptive network are combined in this paper Compare An improved multi-level adaptive network cascaded Hidden Markov Model system which can make explicit use of accent information is proposed and applied to a speech lacking accent Mandarin speech recognition task composed of four regional accents. After a series of discriminative training and adaptation, the system was significantly better than the baseline Accent Independent Deep Neural Network cascade system with an absolute decrease in word error rate of 0.8% to 1.5% (6% to 9% relative) .