Supplementary Material for: Voice Signatures Of Momentary Psychological Stress In Real-Life Environments: Results From The Colive Voice Study

Introduction
Voice is hypothesized to be modulated by stress, and thus could be used as a potential stress detection and monitoring solution. In the literature, vocal biomarkers for stress have mostly been developed on experimental data, with limited samples. Therefore, this study aims to present insights into the effect of momentary psychological stress on voice in real-life recordings, across different languages, genders, and vocal tasks.

Methods
Participants from the Colive Voice study reported their stress level on a 1 to 5 Likert scale. Two tasks were performed: a text-reading task and an A-vowel phonation. We analyzed the data cross-sectionally. We extracted vocal features with the DisVoice library and performed ordinary least squares regression models to evaluate the association of vocal features with stress. Models were stratified by gender and language (French/English) and controlled for age, smoking status, alcohol consumption, the presence of chronic disease, education level, mother tongue, well-being, fatigue, and depression. Benjamini-Hochberg correction was applied to control for multiple testing.

Results
We analyzed a sample of 4155 participants, 2011 in French (1621 women, 390 men) and 2144 in English (1105 women, 1039 men). In the text reading task, we found that stress was associated with two articulatory features for English-speaking women. Among French-speaking women, higher stress was linked with lower pitch and higher shimmer. The duration of pauses and one glottal feature were also associated with stress. In the A-vowel phonation task, pitch and the variability of the pitch perturbation quotient were lower with stress in English-speaking men. French-speaking women had increased voice intensity and loudness with stress.

Conclusion
We were able to confirm the association of momentary psychological stress with various vocal features in real-life settings, but not across languages, vocal tasks, or gender. Future research should include longitudinal studies to investigate the potential of using voice as an intra-individual monitoring biomarker for stress.

## 引言
语音被认为受压力调控，因此可作为潜在的压力检测与监测方案。现有研究中，用于压力识别的语音生物标志物大多基于实验数据开发，且样本量有限。因此，本研究旨在探究瞬时心理压力对真实场景录制的语音的影响，覆盖不同语言、性别与语音任务类型。

## 方法
本研究依托Colive Voice研究的受试者数据，所有参与者通过1至5级李克特（Likert）量表自评压力水平。实验包含两项语音任务：文本朗读任务与元音/a/发音任务。本研究采用横断面分析方法，使用DisVoice工具库提取语音特征，并构建普通最小二乘回归模型，以评估语音特征与压力水平的关联。模型按性别与语言（法语/英语）进行分层，并控制年龄、吸烟状态、饮酒情况、慢性病患病情况、受教育水平、母语、健康状况、疲劳程度与抑郁状态等混杂因素。同时采用本杰米尼-霍奇伯格（Benjamini-Hochberg）校正控制多重检验偏差。

## 结果
本研究最终纳入4155名受试者的数据分析，其中法语组2011人（女性1621名、男性390名），英语组2144人（女性1105名、男性1039名）。在文本朗读任务中，英语女性群体的压力水平与两项发音特征存在显著关联；法语女性群体中，较高压力水平与更低的基频（pitch）、更高的振幅微扰（shimmer）相关。此外，停顿时长与一项声门特征也与压力水平存在关联。在元音/a/发音任务中，英语男性群体的压力水平越高，其基频与基频扰动商（pitch perturbation quotient）的变异性越低；法语女性群体的压力水平越高，语音强度与响度越高。

## 结论
本研究证实了真实场景下瞬时心理压力与多种语音特征存在关联，但该关联未在不同语言、语音任务或性别群体中保持一致。未来可开展纵向研究，以探究语音作为个体内压力监测生物标志物的应用潜力。