腭裂语音高鼻音等级自动识别算法研究

为了对腭裂语音的高鼻音进行等级区分,本文提出基于声学特征参数分析的腭裂语音高鼻音等级自动识别算法，提取基于香农能量和Mel倒谱系数（Mel Frequency Cepstrum Coefficient MFCC）的S-MFCC作为声学特征参数，结合高斯混合模型（Gaussian Mixture Model GMM）分类器实现对腭裂语音四类高鼻音等级（正常、轻度、中度和重度）的自动识别。实验结果表明，提出的自动识别算法取得了较高的高鼻音类别识别率，对四类高鼻音的平均识别率达到79%以上，其中能量与Mel倒谱系数组成的特征参数组取得了85%的平均识别率,优于传统的香农能量算法，具有较高的临床应用价值。

Automatic Hypernasal Detection Based on Acoustic Analysis in Cleft Palate Speech

To detect hypernasal automatically for cleft palate patients, an automatic hypernasal detection algorithm was proposed, based on Shannon energy and Mel frequency cepstrum coefficient acoustic features, combined with Gaussian mixture model classifier. The experiment results showed that the presented method achieved a good performance on the detection of four levels of hypernasal: normal, low-level, moderate-level and high-level. The average classification accuracies for four levels of hypernasal were over 79%. Moreover, the correct recognition accuracy using energy plus Mel frequency cepstrum coefficient feature set reached up to 85%. The classification of hypernasal levels has important clinical applications.