摘要
本文讨论了人工神经网络技术在语音压缩编码方面的应用,提出了一种用Kohonen网络实现语音多脉冲激励分析模型的矢量量化方法。该方法将参数分析和量化编码熔为一体,和传统的先分析、后量化方法相比较,具有许多优良的特性,如全并行处理、过程简化等。本文针对语音多脉冲激励模型,提出了量化网络的结构和学习规则,并将此方法和传统方法进行了比较。最后对网络的压缩性能进行了计算机模拟,结果表明应用人工神经网络进行语音信源的压缩是切实可行的。
In this paper, the application of neural network in the speech compression encoding is discussed. A vector qoantizer based on the artificial neural network is provided which can be used in the quanti-zation of multi-pulse excitation analysis model of the speech signal. The quantizing network is somewhat similar to the Kohonen's net. It performs the parameter analysis and quantizing process together. Comparing with the traditional method which performs analysis first and then quantization, it has some excellent prop erties. We provide the architecture and learning rule of the quantizing network, and compare it with the traditional method in the implementation. Finally, we simulate the quantizing network with practical speech signal. The experimental results show that our scheme is feasible.
出处
《通信学报》
EI
CSCD
北大核心
1992年第5期3-10,共8页
Journal on Communications
关键词
语音
夭量量化
神经网络
脉冲激励
Kohonen self-organizing feature mapping, Speech Compression encoding, Fully vector quan-tization, Parameter analysis.
