基于递归神经网络的语音识别快速解码算法

发布时间：2018-04-16 23:08

本文选题：语音识别 + 递归神经网络　；参考：《电子与信息学报》2017年04期

【摘要】：递归神经网络(Recurrent Neural Network,RNN)如今已经广泛用于自动语音识别(Automatic Speech Recognition,ASR)的声学建模。虽然其较传统的声学建模方法有很大优势,但相对较高的计算复杂度限制了这种神经网络的应用,特别是在实时应用场景中。由于递归神经网络采用的输入特征通常有较长的上下文,因此利用重叠信息来同时降低声学后验和令牌传递的时间复杂度成为可能。该文介绍了一种新的解码器结构,通过有规律抛弃存在重叠的帧来获得解码过程中的计算开销降低。特别地,这种方法可以直接用于原始的递归神经网络模型,只需对隐马尔可夫模型(Hidden Markov Model,HMM)结构做小的变动,这使得这种方法具有很高的灵活性。该文以时延神经网络为例验证了所提出的方法,证明该方法能够在精度损失相对较小的情况下取得2~4倍的加速比。
[Abstract]:Recurrent Neural Network (RNN) has been widely used in acoustic modeling of automatic Speech recognition (ASR).Although it has a great advantage over the traditional acoustic modeling method, the relatively high computational complexity limits the application of this neural network, especially in real-time application scenarios.Since the input features used in recurrent neural networks usually have a long context it is possible to reduce the time complexity of acoustic posteriori and token passing by using overlapping information at the same time.In this paper, a new decoder structure is introduced, which can reduce the computation cost of decoding by regularly discarding overlapped frames.In particular, this method can be directly applied to the original recurrent neural network model, with small changes in the Hidden Markov Model (HMMs) structure, which makes this method highly flexible.In this paper, the delay neural network is taken as an example to verify the proposed method, and it is proved that the proposed method can achieve a speedup ratio of 2 to 4 times when the precision loss is relatively small.
【作者单位】：中国科学院声学研究所语言声学与内容理解重点实验室;中国科学院大学;中国科学院新疆理化技术研究所新疆民族语音语言信息处理实验室;
【基金】：国家自然科学基金(U1536117,11590770-4) 国家重点研发计划重点专项(2016YFB0801200,2016YFB0801203) 新疆维吾尔自治区科技重大专项(2016A03007-1)~~
【分类号】：TP183;TN912.34

【相似文献】