美国 Scicon R&D 多通道数据采集系统 X516

摘要 · 中文整理：本研究运用机器学习方法，考察音段时长、谱矩与离散余弦变换（DCT）系数在区分奥罗莫语擦音单辅音与长辅音（叠辅音）方面的作用。十八位西部方言的奥罗莫语母语者朗读了包含元音间擦音的语料。研究从语料中点位置提取音段时长、经 bark 变换的谱矩以及前六个 DCT 系数等声学特征，并用支持向量机、随机森林与多层感知机神经网络对这些音进行分类。结果显示，音段时长是区分单辅音与长辅音最稳定的特征，DCT 系数的表现略优于谱矩；将时长与谱矩结合可获得最高的分类准确率，而仅使用非时间性特征则误判更多。

原文摘要：Using machine learning, this study investigates how segment duration, spectral moments, and DCT coefficients help differentiating singleton from geminate in Oromo fricatives. Eighteen native Oromo speakers from the western dialect produced speech data consisting of the fricatives in an intervocalic position. Acoustic features, including segment duration, bark-transformed spectral moments, and the first six DCT coefficients, were extracted from the midpoint of the speech data. Support Vector Machine, Random Forest, and Multilayer Perceptron Neural Network classified the sounds. Results reveal that segment duration is the most consistent feature for distinguishing singleton and geminate, with DCT coefficients slightly outperforming spectral moments. The highest classification accuracy could be achieved by combining duration and spectral moments, but non-temporal features result in more errors.

关键词：奥罗莫语、擦音、机器学习、离散余弦变换系数

研究目标：解决奥罗莫语擦音长短对立声学描写不足的问题：此前关于奥罗莫语擦音的声学研究极少，且未系统报告长短（单辅音与叠辅音）对立的声学线索，亦未评估各类时间性与非时间性声学特征在区分长短时的相对贡献。

主要结论：音段时长是区分奥罗莫语擦音单辅音与长辅音最稳定、最可靠的声学特征；非时间性特征中，DCT 系数的区分能力略优于谱矩，但单独使用非时间性特征会带来更多误判；将时长与谱矩组合后可获得最高的分类准确率，说明长短对立在奥罗莫语擦音中主要由时长承载，谱形特征起补充增强作用。

创新点：首次将支持向量机、随机森林与多层感知机神经网络三种机器学习分类器应用于按音长对立划分擦音类别的任务，并在同一数据集上系统比较了时长、bark 谱矩与 DCT 系数三类特征及其组合的分类表现，同时给出特征重要度排序，为奥罗莫语擦音提供了此前缺失的声学与分类证据。

后续建议：增加每个目标词的语料条数与词表规模，现有设计每音仅一个词条，限制了结果的可推广性；扩展到东部与南部方言并纳入该方言区特有的清软腭擦音，检验结论的跨方言稳定性；补充发音生理数据（如电子腭位或气流气压通道），验证长辅音是否伴随更强的气流与更窄的收紧度；引入动态谱特征与多时间点采样，替代仅取摩擦段中点的静态取样方式；加入深度学习模型与跨说话人交叉验证，评估模型在未见说话人上的泛化能力

APA 引文：Negesse, F. (2025). Classifying Oromo fricatives with machine learning. Estudios de Fonetica Experimental, 34, 151-168. https://doi.org/10.1344/efe-2025-34-151-168

仪器引用：The speakers were recorded while saying the words presented in a random order in a Keynote on MacBook Air (version 2019). The recording took place in two sessions with PCquirer (Sicon R & D Inc.) in the Phonetics Lab of the Department of Linguistics.

摘要 · 中文整理：本文首次详细探索性地描写了刚果西南部班图方言群萨卡塔语中唇软腭口塞音的空气动力学特征。数据在蒙斯大学语音实验室采集，发音人为 3 名中部萨卡塔语母语者。文章给出了唇软腭塞音与普通双唇口塞音的对比数据，讨论了相关变量的描述统计结果，并针对整套数据中专门划分的子集给出多元方差分析结果，以考察语料中的变异来源。结果显示，萨卡塔语唇软腭塞音在时长、气流与气压模式上均不同于普通双唇塞音；清唇软腭塞音的气压与气流数值表明其舌根/喉部下降幅度比浊音对应项更为显著。文章还梳理并讨论了本研究结果与已有类型学文献之间的一致与不一致之处。

原文摘要：The present contribution represents the first in-detail exploratory account of the aerodynamics of labial-velar oral stops in Sakata, a Bantu dialect cluster of southwestern Congo. Data collection took place at the phonetics laboratory facilities of Universite de Mons with three speakers of central Sakata. Comparative data of labial-velar and plain bilabial oral stops are presented and analysed. Descriptive statistics of the relevant variables are discussed. Given each group of variables, MANOVA results are presented for specially tailored subsets of the whole dataset to investigate variance in the corpus. Sakata labial-velar stops are shown to differ from plain bilabials for duration, air-flow, and pressure patterns. Voiceless labial-velar stops exhibit pressure and airflow values consistent with a more prominent lowering of the tongue root / larynx than their voiced counterparts. Matches and mismatches with the available typological literature are also delineated and discussed.

关键词：空气动力学、唇软腭音、班图语言、语音通信、语音记录与描写

研究目标：解决唇-软腭双重调音塞音缺乏细致器械化空气动力学描写的问题：既有文献多凭印象或少量个案假定这类塞音在两个闭塞之间存在口腔气流稀薄化，但萨卡塔语此前只有声学描写，清浊两类唇-软腭塞音与普通双唇塞音在时长、气流方向与口内气压模式上究竟如何区分，尚属未知。

主要结论：萨卡塔语唇-软腭塞音与普通双唇塞音在时长、口腔气流与口内气压三方面均存在显著差异：唇-软腭塞音在首次闭塞与除阻处气流量明显更大，除阻时出现正向气流突增，说明存在肺部呼气气流机制；双唇塞音在除阻时气流接近零而口内压更高。清音唇-软腭塞音比浊音对应项的闭塞期更长、T3 处口内压更低，反映出更显著的舌根/喉部下降与口腔气体稀薄化；清浊对立在低频能量（浊音杠）上并无清晰的谱学线索，提示其对立更依赖喉部下降等发音机制。

创新点：首次以 PCquirer X516 多通道气流气压系统对萨卡塔语唇-软腭塞音做了声学与三通道空气动力学同步采集，并提出以口内压与口腔气流曲线为依据的 T1—T4 四点标注框架，把双重调音塞音的时序结构与气流气压事件直接挂钩；同时采用先 PCA 降维再 MANOVA 的统计路径处理高共线性的时长变量群。

后续建议：扩大发音人样本量与方言覆盖面，现有三位旅居欧洲的发音人难以代表萨卡塔语整体，后续应在刚果当地开展田野采集；纳入唇-软腭鼻音及鼻音前接语境，检验文中关于前接鼻音是否同为双重调音的推测；补充电子声门仪或光电声门仪等喉部观测手段，直接验证舌根/喉部下降的假设，而非仅由气压推断；引入线性混合效应模型以发音人为随机效应，替代或补充 MANOVA，提升小样本条件下统计推断的稳健性；同步采集鼻腔气流并纳入分析，现有研究已记录该通道但未用于本次分析，可进一步刻画鼻音-口音过渡

APA 引文：Maselli, L., & Delvaux, V. (2024). Aerodynamics of Sakata labial-velar oral stops. In Proceedings of Interspeech 2024 (pp. 3140-3144). ISCA. https://doi.org/10.21437/Interspeech.2024-525

仪器引用：Aerodynamic data were collected with PcQuirer 516 (Scicon R and D Inc.). This allowed us to simultaneously record acoustic traces and three channels of aerodynamic data (mouth pressure, oral airflow, nasal airflow). The audio signal was digitised and sampled at 12 kHz, dc channels were sampled at 1 kHz. Mouth pressure was obtained by dint of a catheter introduced at the side of the mouth, bent behind the second molars, and connected to a pressure transducer. Volume flow from the mouth was collected with a Rothenberg mask, with a pressure tube inserted in one of the mask's outlet holes. Pressure and airflow signals were low-pass filtered at 50 Hz.