基于正弦模型的语音识别时频特征.pdfVIP

5 1 解放 军理 工大学 学 报( 自然 科学 版) Vol. 5 No . 1 2004 2 Jour nal of PLA Univer sity of S ien e and T e hnology Feb. 2004 : 1009-3443( 2004) 01-0022-04 基于正弦模型的语音识别时频特征 邢艳玲, 　杨吉斌, 　张雄伟 ( , 2 10007) : 为改善语音识别系统的性能, 采用时频分布参数来 述语音特征。由于时频分布参数考虑到语音信 号内在的非平稳特性, 因此能够更准确地 述语音信号的时频特性。对基于正弦模型的多种时频参数( 能量 谱和幅度加权瞬时频谱) 进行了比较, 并在基于隐马尔可夫模型的连接词语音识别系统中进行了实验仿真。 结果表明, 单独采用时频分布参数作为ASR 的前端特征并不能改善识别率; 而采用标准ASR 特征和能量谱 时频特征的联合前端特征, 可以有效地改善语音识别系统的识别效果。 : 语音识别; 语音前端特征; 时频分布; 正弦模型; 能量谱 : T N919. 81 : A - T ime frequency Feat ures for Speech Recognit ion Based on Sinusoidal Speech Model X I N G Yan-ling , YA N G J i-bin, ZH A N G X iong -w ei ( Institute of Communi ations Engineering , PLA Univ . of S i. T e h. , Nanjing 210007, China) Abstract: T he use of time-frequen y distribution as features for automati spee h re ognition ( ASR) is in- tended to improve the performan e of ASR. Be ause of the inherent non-stationary hara teristi of spee h , - . - , signal time frequen y distribution an des r ibe the signal more pre isely T ime frequen y features su h as ener gy spe trum and amplitude w eighted short-time average of the instantaneous frequen y, are om- pared based on the sinusoidal spee h model, and some experiments are arried out in onne ted digits . - re ognition system based on Hidden M ar kov M odels T he results indi ate that time frequen y features as - , fr ont end ant improve the performan e of ASR solely w hile the feature w hi h ombines standard ASR featur es and time-frequen y features based on ener gy spe trum an improve the performan