古典唱法中喉位稳定性、共鸣与声音开放度的声学结构及其与基于机器学习的教学评分的关联
Acoustic Structure of Laryngeal-Position Stability, Resonance, and Vocal Openness in Classical Singing and Their Associations with Machine-Learning-based Pedagogical Ratings.
文献信息
| PMID | 42919718 |
|---|---|
| 原文 | 在 PubMed 查看原文 ↗ |
| 发表日期 | 2026 |
| 作者 | Zhenyi Hou |
| 作者单位 | School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China; Hujiang College, University of Shanghai for Science and Technology, Shanghai 200093, China. |
| 期刊 | Journal of voice : official journal of the Voice Foundation |
| SCI 分区 | Q1 |
| IF | 2.4 |
| 研究类型 | AI/ML · 临床 |
| 所属专科 | 咽喉科 |
中文摘要
背景： 古典声乐技术中的关键概念，包括喉位稳定性、共鸣和声音开放度，是声乐教学的核心，但在很大程度上仍依赖听觉判断。它们与可重复的声学模式之间的关系仍未被充分理解。
目的： 本研究使用机器学习模型，探讨可解释的声学特征是否能与专家对这三个教学维度的评分相关联，并考察联合声乐与核心肌群稳定性（C-VaC）热身后的声学变化。
方法： 分析了来自12名受过训练的女性歌手的3000个音频样本数据集，包括C-VaC热身前后配对的录音。提取声学特征，并分别开发支持向量机（SVM）模型，以预测1-5序数量表上的专家评分。使用受试者独立交叉验证评估模型性能。
结果： 在10折交叉验证下，模型的总体宏F1分数为0.74，准确率为0.72。热身之后最一致的声学改善出现在中声区，包括嗓音扰动降低和倒谱峰值突出度增加。
结论： 多维声学特征有助于对古典声乐教学法中关键概念的专家一致评分进行建模。该方法可用定量信息补充传统听觉评估，但仍需在更大、更多样化的歌手人群中进行进一步验证。
英文摘要
BACKGROUND: Key concepts in classical vocal technique, including laryngeal-position stability, resonance, and vocal openness, are central to voice teaching but remain largely dependent on auditory judgment. Their relationships with reproducible acoustic patterns remain insufficiently understood.
OBJECTIVE: This study investigated whether interpretable acoustic features could be associated with expert ratings of these three pedagogical dimensions using machine-learning models and examined acoustic changes following Combined Vocal and Core Muscle Stability (C-VaC) warm-up.
METHODS: A dataset of 3000 audio samples from 12 trained female singers, including paired recordings before and after C-VaC warm-up, was analyzed. Acoustic features were extracted, and separate support vector machine (SVM) models were developed to predict expert ratings on a 1-5 ordinal scale. Model performance was evaluated using subject-independent cross-validation.
RESULTS: Under 10-fold cross-validation, the models achieved an overall macro-F1 score of 0.74 and an accuracy of 0.72. The most consistent acoustic improvements after warm-up were observed in the middle register, including reduced vocal perturbation and increased cepstral peak prominence.
CONCLUSIONS: Multidimensional acoustic features can help model expert-aligned ratings of key concepts in classical vocal pedagogy. This approach may complement traditional auditory evaluation with quantitative information, although further validation in larger and more diverse singer populations is needed.