基于AWS_VFR的语音特征提取方法

Variable frame rate based on adaptive weighted-sum for speech feature extraction

针对语音识别中固定帧率特征提取方法没有充分考虑语音频谱变化特性、噪声鲁棒性差的问题,提出了一种基于自适应加权和的变帧率方法用于特征提取,并在固定音频检索系统中进行实验,在信噪比为20 dB的情况下,与固定帧率的特征提取方法相比,系统检出率提高了近4%。实验表明,该方法在降低噪声影响,提高固定音频检索性能方面是有效的。

Given the problems of fixed frame rate feature extraction method which did not fully consider the changes in charac- teristics of speech spectrum, and had a poor noise robustness, this paper proposes a variable frame rate method based on adap- tive weighted-sum for speech feature extraction, and experiments are conducted with a specific audio retrieval system. Com- pared to the fixed frame rate feature extraction method, under the signal noise ratio of 20 dB, the system detection rate increases nearly 4%. Experimental results show that this method is effective to reduce noise, and to improve the performance of fixed au- dio retrieval.