考虑帧间信息的语音带宽扩展

• 作 者: 王迎雪 赵胜辉 匡镜明
• 机 构: 北京理工大学信息与电子学院,卡内基梅隆大学计算机与工程学院,
• 出 处: 声学学报
• 发表日期: 2017年03期
• 关键词: 帧间 玻尔兹曼机 高斯混合模型 特征参数 分段线性函数 统计特性 非线性关系 方法性能 距离值 条件受限

语音带宽扩展是为了提高语音质量,利用语音低频和高频之间的相关性重构语音高频的一种技术。高斯混合模型法是语音带宽技术中被广泛应用的一种方法,但是,该方法的映射函数是分段线性函数,且没有考虑语音前后帧的相关信息。因此,提出了一种基于条件受限玻尔兹曼机的方法。该方法利用条件受限玻尔兹曼机提取了语音信号的帧间信息,同时将语音低频、高频特征参数映射为高阶统计特性,深层发掘和模拟了语音低频和高频之间的非线性关系。客观和主观对比测试结果都表明,该方法性能优于传统的高斯混合模型方法。

Voice bandwidth extension is a technique for reconstructing speech high frequencies using the correlation between speech low frequencies and high frequencies to improve speech quality. Gaussian mixture model is widely used in speech bandwidth technology. However, the mapping function of this method is piecewise linear function, and does not consider the relevant information before and after the speech. Therefore, a method based on conditional limited Boltzmann machine is proposed. The method utilizes the limited Boltzmann machine to extract the inter-frame information of the speech signal and maps the low-frequency and high-frequency characteristic parameters of the speech into the high-order statistical properties. The method further explores and simulates the low-frequency and high-frequency non-speech Linear relationship. Both objective and subjective comparison test results show that this method is superior to the traditional Gaussian mixture model.