摘要
研究表明,增强后的语音与纯净语音相比,会存在两种不同类型的畸变:放大畸变和衰减畸变,而放大畸变对语音可懂度的影响较大。传统的语音增强算法大多不能有效提高语音增强后的可懂度,因为这些算法仅使用最小均方误差的方法来限制这两种畸变,从而抑制噪声,提高语音的质量,但忽略了不同的畸变类型对可懂度的影响不同。提出一种基于子空间的提高可懂度的语音增强算法,使用先验信噪比及增益矩阵来判断语音畸变的类型。同时注意到,在估计先验信噪比时会存在估计误差:高估和低估,而高估会产生放大畸变,对可懂度造成较大的影响。先对高估先验信噪比(小于-10 dB)的增益矩阵进行修正,然后再对幅度谱畸变大于0 dB及6.02 dB的语音进行不同的限制。实验表明,所提出的算法能够有效增强语音的可懂度。
Research shows that compared with pure speech,there will be two types of distortion in enhanced speech: the amplifier distortion and the attenuation distortion,and the amplifier distortion has a bigger influence on the speech intelligibility. Most of traditional speech enhancement algorithms cannot effectively improve the intelligibility after the speech being enhanced,because they only use the minimum mean square error method to limit these two kinds of distortions for suppressing noise and improving the quality of the speech,but neglect the different distortions having different effects on intelligibility. This paper proposes a subspace-based speech enhancement algorithm which can improve the intelligibility,it uses priori signal-to-noise ratio( SNR) and gain matrix to judge the type of speech distortion. At the same time,it is noticed that when estimating the priori SNR,there is the estimation error: overestimation and underestimation. The overestimation will produce amplifier distortion,and has a big influence on the intelligibility. We first correct the gain matrix of the overestimated priori SNR( less than- 10 dB),and then make different restrictions on the speeches with amplitude spectrum distortion greater than 0 dB and 6. 02 dB separately. Experimental results show that the proposed algorithm can effectively enhance the intelligibility of the speech.
出处
《计算机应用与软件》
CSCD
2016年第10期145-147,180,共4页
Computer Applications and Software
基金
国家自然科学基金项目(61371193)
