基于回声状态网络的音频频带扩展方法  

Audio Bandwidth Extension Method Based on Echo State Network

作　　者：刘鑫[1] 鲍长春[1] 

机构地区：[1]北京工业大学电子信息与控制工程学院,北京100124

出　　处：《电子学报》2016年第11期2758-2766,共9页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61072089;No.61471014)

摘　　要：宽带音频通信系统对传输信号有效带宽的限制会降低重建音频的主观质量和自然程度.本文提出了一种基于回声状态网络的宽带向超宽带音频盲目式频带扩展方法.该方法借助回声状态网络来模拟音频信号高低频频谱参数间的映射关系,并依据网络模型中的时延递归结构连续更新系统状态来近似描述音频特征的时域演变过程,有效地估计了高频成分的频谱包络.同时,结合频谱复制方法得到的高频频谱细节,该方法实现了宽带向超宽带音频的有效扩展.测试结果表明,本文所提方法提升了宽带音频的听觉质量;对于多数测试数据,该方法在静态和动态失真方面获得了优于高斯混合模型扩展方法的扩展性能.The bandwidth limitation in wideband audio communication systems degrades the subjective quality and naturalness of the reproduced signals. In this paper, a wideband to super-wideband audio bandwidth extension method was proposed by using echo state network. The echo state network is adopted to model the mapping function between the low-and high-frequency spectral coefficients of audio signals, and the temporal evolution of audio features is represented by continu- ously state updating on the basis of the recursive structure in the network, for effectively estimating the high-frequency spec- tral envelope. By combining the high-frequency fine spectrum extended by spectral translation, the proposed method can ef- fectively extend the bandwidth of wideband audio to super-wideband. Evaluation results show that the proposed method up- grades the auditory quality of wideband audio, and gains better extension performance than the Gaussian mixture model- based bandwidth extension method in terms of both static and dynamic distortions for most test data.

关 键 词：音频编码 音频频带扩展 回声状态网络 频谱复制 

分 类 号：TN912.3[电子电信—通信与信息系统]