语音识别特征提取中对特征方法的对比

Comparison of feature methods in feature extraction for speech recognition

人工智能概念的提出,让语音识别迎来了新的生机。随着相关知识与技能的飞速发展,神经网络带动了语音识别领域相关知识的革新。文章使用语音识别中常见的LPCC特征、MFCC特征和PLP特征对同一段语音进行特征提取,通过特征图像化可以直观展示其特征的优劣势。其中,LPCC特征对频谱包络变化较为敏感;MFCC特征具有较好语音信号的短时频谱,对信号的语音干扰和音量变化等抗干扰能力较好,但高频细节不够清晰;PLP特征具有较好的鲁棒性,对信号的语音干扰和音量变化等有很好的抗干扰能力,且对高频部分的细节信息表示更为准确。

The introduction of the concept of artificial intelligence has ushered in new vitality for speech recognition.With the rapid development of related knowledge and skills,neural networks have driven the innovation of relevant knowledge in the field of speech recognition.This article uses common LPCC features,MFCC features,and PLP features in speech recognition to extract features from the same segment of speech.Through feature visualization,the advantages and disadvantages of these features can be visually displayed.Among them,LPCC features are more sensitive to changes in spectral envelope.MFCC features have a good short-term spectrum of speech signals,and have good anti-interference ability against speech interference and volume changes,but high-frequency details are not clear enough.PLP features have good robustness and have good anti-interference ability against speech interference and volume changes in signals,and are more accurate in representing detailed information in high-frequency parts.