CAJ | 학술논문

针对非结构噪声难以去除的问题，基于字典训练和稀疏表示提出一种无监督语音增强算法。该算法通过构造过完备字典并使用带噪语音样本对其进行训练来实现。首先指出K-奇异值分解算法( K-SVD)存在的不足并提出一种新的改进的字典训练算法：K-双边随机投影算法( K-BRP)；然后使用K-BRP算法不断更新字典矩阵和相应的增益系数矩阵，从被非结构化噪声所污染的带噪语音中提取出结构性强的纯净语音。大量实验结果表明，由于训练样本考虑到了语音信号的时频域局部结构特征，所提算法能够很好地消除随机噪声，并且在低信噪比情况下仍然能够保持较高的语音质量和可懂度。
침대비결구조성난이거제적문제，기우자전훈련화희소표시제출일충무감독어음증강산법。해산법통과구조과완비자전병사용대조어음양본대기진행훈련래실현。수선지출K-기이치분해산법( K-SVD)존재적불족병제출일충신적개진적자전훈련산법：K-쌍변수궤투영산법( K-BRP)；연후사용K-BRP산법불단경신자전구진화상응적증익계수구진，종피비결구화조성소오염적대조어음중제취출결구성강적순정어음。대량실험결과표명，유우훈련양본고필도료어음신호적시빈역국부결구특정，소제산법능구흔호지소제수궤조성，병차재저신조비정황하잉연능구보지교고적어음질량화가동도。
To solve the difficulty in enhancing the speech contaminated by unstructured noise, an unsupervised speech enhancement algorithm based on dictionary training and over-completely representation was proposed. In the enhancement stage, the technique alternated between sparse coding of the gain matrix and updating the dictionary atoms by using the K-Bilateral Random Projection ( K-BRP) algorithm which is a faster modification of the K-Singular Value Decomposition ( K-SVD) algorithm. In this way, clean speech was extracted from the noisy speech. Extensive experimental results show that the proposed algorithm achieves better performance in terms of speech quality and speech intelligibility even in low Signal-to-Noise Ratio ( SNR) condition by considering the local time-frequency characteristics of speech.