用于说话人无关嗓音病理检测的临床Praat参数与倒谱表示：年龄与发声障碍严重程度的影响

文献信息

中文摘要

目的： 在说话人无关评估条件下，使用完整的萨尔布吕肯嗓音数据库持续元音语料，确定经临床验证的Praat参数能否区分病理嗓音与健康嗓音；比较其性能与高维倒谱表示及通过自监督训练的语音基础模型的差异；并探讨判别能力如何取决于发声障碍严重程度以及组间年龄差异。
研究设计： 对公开嗓音数据库进行回顾性分析。机器学习任务被设定为二分类：健康嗓音与病理嗓音。
方法： 分析了来自1180名说话者（健康组年龄27.9 ± 11.7岁，病理组年龄50.7 ± 15.4岁）的1288次会话中的全部15,442条持续元音录音。共提取78个特征：通过Praat提取14个临床参数，通过librosa提取64个倒谱、频谱和能量描述符。在会话水平上使用Mann-Whitney U检验组间差异，以秩双列相关量化，并经Holm-Bonferroni校正。采用按说话者分组的五折嵌套交叉验证比较六种分类器。声学改变严重程度通过声学嗓音质量指数（AVQI）估计，该指数不能替代感知评估；年龄通过按年龄和性别将患者与对照匹配以及对训练数据重新加权进行控制。
结果： 11个参数中有9个达到中等效应量，以jitter PPQ5（r = -0.414）和CPPS（r = +0.382）为首。60%的病理会话其声学嗓音质量处于健康范围内。SVM（0.775 ± 0.019）和XGBoost（0.770 ± 0.013）达到最高的录音水平AUC；适配的wav2vec 2.0达到0.726 ± 0.015。14个Praat参数与52个倒谱系数表现相似（AUC差异 +0.014，95% CI -0.002至+0.030）；既未证明等效性，也未证明优越性。控制年龄后，XGBoost达到会话水平AUC 0.721（95% CI 0.669至0.769）。判别能力在由AVQI定义的声学改变分层中逐步增加：健康范围内为0.60，高于健康范围为0.89至0.99。年龄匹配后，jitter、shimmer和CPPS保持其效应量，而f0变异性则未保持。
结论： 持续元音的声学分析可识别具有可测量声学改变的发声障碍，但不能识别声学上轻微的障碍；14个可解释参数的表现与规模近四倍的倒谱表示相似。在SVD上报告的判别能力应结合组间年龄差异和样本的严重程度构成来解读。

英文摘要

OBJECTIVES: To determine whether clinically validated Praat parameters discriminate pathological from healthy voice under speaker-independent evaluation of the complete Saarbrücken Voice Database corpus of sustained vowels, how their performance compares with a high-dimensional cepstral representation and a speech foundation model trained through self-supervision, and how discrimination depends on dysphonia severity and on the age difference between groups.
STUDY DESIGN: Retrospective analysis of a publicly available voice database. The machine learning task was formulated as binary classification: healthy versus pathological voice.
METHODS: All 15,442 sustained vowel recordings from 1288 sessions of 1180 speakers (age 27.9 ± 11.7 years in healthy and 50.7 ± 15.4 years in pathological speakers) were analyzed. A total of 78 features were extracted: 14 clinical parameters via Praat and 64 cepstral, spectral, and energy descriptors via librosa. Group differences were tested at the session level with Mann-Whitney U, quantified by rank-biserial correlation, and corrected by Holm-Bonferroni. Six classifiers were compared using nested cross-validation with five folds grouped by speaker. The severity of acoustic alteration was estimated with the Acoustic Voice Quality Index (AVQI), which does not replace perceptual evaluation, and age was controlled by matching patients to controls on age and sex and by reweighting the training data.
RESULTS: Nine of 11 parameters reached medium effect size, led by jitter PPQ5 (r = -0.414) and CPPS (r = +0.382). Sixty percent of pathological sessions had acoustic voice quality within the healthy range. The SVM (0.775 ± 0.019) and XGBoost (0.770 ± 0.013) reached the highest recording-level AUC; adapted wav2vec 2.0 reached 0.726 ± 0.015. The 14 Praat parameters and the 52 cepstral coefficients performed similarly (AUC difference +0.014, 95% CI -0.002 to +0.030); neither equivalence nor superiority was demonstrated. With age controlled, XGBoost reached a session-level AUC of 0.721 (95% CI 0.669 to 0.769). Discrimination increased progressively across strata of acoustic alteration defined by the AVQI: 0.60 within the healthy range and 0.89 to 0.99 above it. Jitter, shimmer, and CPPS kept their effect sizes after age matching, whereas f0 variability did not.
CONCLUSIONS: Acoustic analysis of sustained vowels identifies dysphonia with measurable acoustic alteration but not acoustically mild disorders, and 14 interpretable parameters performed similarly to a cepstral representation nearly four times larger. Discrimination reported on the SVD should be interpreted in the light of the age difference between groups and the severity composition of the sample.