语音识别关键技术研究

采用隐马尔可夫模型(HMM)进行语音声学建模是大词汇连续语音识别取得突破性进展最主要的原因之一,HMM本身依赖的某些不合理建模假设和不具有区分性的训练算法正在成为制约语音识别系统未来发展的瓶颈。神经网络依靠权能够进行长时间记忆和知识存储,但对于输入模式的瞬时响应的记忆能力比较差。采用混合HMM/ANN模型对HMM的一些不尽合理的建模假设和训练算法进行了革新。混合模型用神经网络非参数概率模型代替高斯混合器(GM)计算HMM的状态所需要的观测概率。另外对神经网络的结构进行了优化,取得了很好的效果。

Key Technology Research for Speech Recognition

Because of the application of the Hidden Markov Model(HMM) in acoustic modeling,a significant breakthrough has been made in recognizing continuous speech with a large glossary.However,some unreasonable hypotheses for acoustic modeling and the unclassified training algorithm on which the HMM based form a bottleneck,restricting the further improvement in speech recognition.The Artificial Neural Network(ANN) techniques can be adopted as an alternative modeling paradigm.By means of the weight values of the network connections,neural networks can steadily store the knowledge acquired from the training process.But they possess a weak memory,not being suitable to store the instantaneous response to various input modes.To overcome the flaws of the HMM paradigm,we design a hybrid HMM/ANN model.In this hybrid model,the nonparametric probabilistic model(a BP neural network) is used to substitute the Gauss blender to calculate the observed probability which is necessary for computing the states of the HMM model.Besides,we optimize the structure of the network,and experiments show that the hybrid model has a good performance in speech recognition.