新神经网络结构及其在数码语音识别中的应用

New neural network architecture with application in mandarin digit speech recognition

为了提高人工神经网络处理动态信号能力 ,在时延神经网络 ( TDNN )和卷积神经网络 ( CNN)的基础上 ,针对孤立音节的特点 ,提出了一个新的网络结构 ,研究了其学习算法。新网络在进一步改进后用于汉语孤立数码语音识别 ,对特定人和非特定人任务 ,分别达到了 97.7%和 95 .6%的正确识别率 (无拒识 ) ,其性能远远高于多层前向感知机( ML P)和时延神经网络 ,与传统的隐马尔科夫模型 ( HMM)方法是可以相比的。

The ability of neural networks to deal with time dynamic signal was improved with a new neural network architecture specializing in syllable recognition based on the time delay neural network (TDNN) and the convolutional neural network. After tuning, the new network achieves 97.7% and 95.6% correct recognition accuracy without rejection, when applied to speaker dependent and speaker independent isolated mandarin digit recognition. Such performance is much better than those of Multilayer Perceptrons and TDNN and is comparable to the much more popular hidden Markov model methodology.