论文部分内容阅读
提出了一种基于机器学习的耳语音可懂度增强方法.该方法利用已经训练好的2类支持向量机来估计一个二元时频掩蔽值,进而合成增强后的耳语音.输入支持向量机的特征向量GFCCs是基于听觉外周模型进行提取的,具有噪声鲁棒特性.在增强仿真实验中,将该算法同传统语音增强算法进行语音可懂度增强性能比较.客观评价和主观听力实验结果均表明,所提出的方法能有效提高含噪耳语音的听觉可懂度;相比谱减法和log-MMSE方法在低信噪比时无法提高语音可懂度,该方法在低信噪比时仍可有效提高含噪耳语音的听觉可懂度.此外,含噪耳语音通过所提出的方法进行增强后,其可懂度比未增强时明显提高.
This paper proposes a machine learning method based on ear language speech intelligibility enhancement.This method uses the trained two kinds of support vector machines to estimate a binary time-frequency mask value and then synthesizes the enhanced ear speech. Input Support Vector Machine The feature vector GFCCs is extracted based on the perceptual peripheral model and has the characteristics of robustness to noise.Compared with the traditional speech enhancement algorithm, the speech enhancement performance of GFCCs is enhanced compared with the traditional speech enhancement algorithm The results show that the proposed method can effectively improve the audibility of noisy white speech. Compared with spectral subtraction and log-MMSE methods, the speech intelligibility can not be improved at low signal-to-noise ratio. Which can effectively improve the auditory intelligibility of noisy white speech.In addition, the noisy speech improves significantly by the proposed method, and its speech intelligibility is obviously improved when it is not enhanced.