基于语音存在概率和听觉掩蔽特性的语音增强算法

Speech enhancement based on speech-presence probability and auditory masking property

低信噪比下,谱减语音增强法中一直存在的去噪度、残留的音乐噪声和语音畸变度三者间均衡这一关键问题显得尤为突出。为降低噪声对语音通信的干扰,提出了一种适于低信噪比下的语音增强算法。在传统的谱减法基础上,根据噪声的听觉掩蔽阈值自适应调整减参数,利用语音存在概率,对语音、噪声信号估计,避免低信噪比下端点检测(VAD)的不准确,有更强的鲁棒性。对算法进行了客观和主观测试,结果表明:相对于传统的谱减法,在几乎不损伤语音清晰度的前提下该算法能更好地抑制残留噪声和背景噪声,特别是对低信噪比和非平稳噪声干扰的语音信号,效果更加明显。

In low Signal to Noise Ratio （SNR） environment, the trade-off among the amount of noise reduction, the level of musical residual noise and the speech distortion was the key problem of spectral subtraction speech enhancement. An improved speech enhancement algorithm for low SNR was developed in order to decrease the interference of noise on pure speech. This algorithm was based on the traditional spectral subtraction, and the subtraction parameter was self-adaptively adjusted in the light of the masking properties of human auditory. The speech-presence probability was integrated to estimate the signal and noise, which would avoid the inaccuracy of utilizing Voice Activity Detection （VAD）, and increase the robustness. The experimental results demonstrate that this algorithm is effective to reduce the residual noise and the background noise, but not to distort the speech articulation compared to other modified spectral subtraction algorithms, especially for the low SNR noisy speech signal and that being degraded by non-stationary noise.