Amplitude-based speech enhancement with nonnegative matrix factorization for asynchronous distributed recording
Amplitude-based speech enhancement with nonnegative matrix factorization for asynchronous distributed recording
复制标题
DOI：
10.1109/iwaenc.2014.6954007
复制
发表时间：
2014-11
期刊：
影响因子：
--
通讯作者：
Hironobu Chiba;Nobutaka Ono;S. Miyabe;Yu Takahashi;Takeshi Yamada;S. Makino
中科院分区：
文献类型：
--
作者：
Hironobu Chiba;Nobutaka Ono;S. Miyabe;Yu Takahashi;Takeshi Yamada;S. Makino
在本文中，我们研究了基于幅度的语音增强异步分布式记录。在ad-hoc麦克风阵列上下文中，假设不同的异步设备记录语音。结果，相位信息由于采样频率失配而不可靠。针对基于幅度信息代替相位信息的语音增强问题，在时域信道域引入了有监督的非负矩阵分解（NMF）。通过使用单个源观测，预先训练表示从源到每个麦克风的传递函数的增益的基向量。实验结果表明，该方法对采样频率失配具有很好的鲁棒性。
In this paper, we investigate amplitude-based speech enhancement for asynchronous distributed recording. In an ad-hoc microphone array context, it is supposed that different asynchronous devices record speech. As a result, the phase information is unreliable due to sampling frequency mismatch. For speech enhancement based on the amplitude information instead of the phase information, supervised nonnegative matrix factorization (NMF) is introduced in the time-channel domain. The basis vectors, which represents the gain of the transfer function from a source to each microphone, are trained in advance by using single source observation. The experimental evaluations show that this approach is well robust against the sampling frequency mismatch.