Optimization of acoustic feature extraction for language identification
The dimensionalities of the acoustic feature matrix used in language identification studies are often very high.To address the issue of excessive dimensions in acoustic features for language identification,an im-proved method for acoustic feature extraction is proposed.By analyzing the statistical characteristics of some commonly used acoustic features and then combining with their extraction process as well as partial literature arguments,the improved features are obtained by calculating the mean value of each dimension of the features on the frame and then normalizing the vectors to eliminate the influence of the dimensions.This results in the optimization of the traditional feature matrix into a one-dimensional feature vector.Finally,based on the char-acteristics of the improved features,experiments for language identification are conducted using BP neural net-work and Support Vector Machine as the baseline systems on two distinct datasets.The experimental results show that,for the five commonly used acoustic features,the proposed improved method consistently achieves an average identification rate of 95.6%for Dataset1 and 90.2%for Dataset2 under the two models,even with a reduction of 99.8%in data volume,compared to the traditional approaches.In addition,the significant reduction in computational workload achieved by proposed method enhances the adaptability of the algorithm to embedded environments with relatively weak hardware facilities,thereby expanding for the applicability of the algorithm.
Language identificationAcoustic featuresStatistical featuresFeature extraction
国家科技期刊平台
NSTL
万方数据
