基于lpcc和mfcc的藏语语音端点检测算法-模式识别国家重点.pdfVIP

ICMIT’07_208 基于 LPCC 和 MFCC 的藏语语音端点检测算法 李洪波1，于洪志2 中国科学院 自动化研究所 摘要 端点检测是语音识别系统预处理阶段遇到的第一个关键技术。该算法根据藏语元音/ 辅音频谱特性差异，对语音信号分高/低频带后再分别处理的思想，符合藏语语音的清/浊对 立信息分布特点，然后分别提取倒谱系数作为端点检测的特征，因为倒谱特征所含的信息比 其他参数多，能较好地表征语音信号，语音质量好，识别正确率高；检测时采取自适应噪声 参数估计，根据倒谱距离轨迹判决起止端点，仿真结果表明了它的优越性。 关键词 端点检测；LPCC；MFCC；安多藏语 Endpoint Detection Algorithm of Tibetan Pronunciation Based on LPCC and MFCC 1 2 Li Hongbo , Yu Hongzhi Institute of Automation Chinese Academy of Sciences Abstract Endpoint detection is the first essential technology which speech recognition system meets in pre-processing stage. This algorithm which based Tibetan vowel/consonant frequency spectrum characteristic, separately is processed again through the pronunciation signal minute high/low-frequency band, conforms to Tibetan pronunciation clear/muddy opposition information distribution characteristic, then separately withdraws but actually is scored the cepstral coefficient to take the endpoint detection characteristic, because the cepstral coefficient actually scores the information which the characteristic contains compared to other parameters many, can attribute the better attribute pronunciation signal, the pronunciation quality is good, the recognition accuracy is high; When examination adopt the auto-adapted noise parameter to estimate that, decided beginning/end vertex according to ceptrum distantce, the simulation result indicated its superiority. Key words speech endpoint detection；Lpc cepstral coefficient ；Mel-frequency cepstral coefficient ；Ando Tibetan 1作者简介：李洪波(1971一)，女，副教授、研究生，主要研究方向：语音识别，语音编码， 计算机应用 2于洪志，教授、博导，研究方向为多文种信息处理、中文信息 项目资助