TY - GEN
T1 - Evaluating Debate Persuasiveness Through Audio Analysis and Regression Techniques
AU - Nott, Gage
AU - Vadapalli, Hima
AU - van der Haar, Dustin
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - In competitive debates, the effectiveness of arguments is often assessed through verbal communication. However, nonverbal biometric factors, such as vocal characteristics, play a crucial yet underexplored role in influencing judges’ perceptions and scores. The proposed study explores the role and significance of nonverbal biometric factors in determining the persuasiveness of arguments in competitive debates. The experimental pipeline includes phases such as data collection, analysis of audio features such as Short-Time Fourier Transforms (STFT) and Mel Spectrograms, and utilization of several machine learning algorithms, including Least Squares Linear Regression, Random Forests (RF), Support Vector Machines (SVM), and a Convolutional Neural Network (CNN) to evaluate the usefulness of nonverbal biometrics in predicting judges’ scores. From the existing IBM Debater dataset of recorded speeches, a subset of 72 speeches across 9 speakers was selected and scored by a team of qualified school-level adjudicators to create the dataset used in these experiments. The preliminary results on the dataset were promising and have provided valuable insights into the challenges and efficacy of various regression techniques in audio-based persuasiveness prediction, highlighting the need for further exploration in this domain.
AB - In competitive debates, the effectiveness of arguments is often assessed through verbal communication. However, nonverbal biometric factors, such as vocal characteristics, play a crucial yet underexplored role in influencing judges’ perceptions and scores. The proposed study explores the role and significance of nonverbal biometric factors in determining the persuasiveness of arguments in competitive debates. The experimental pipeline includes phases such as data collection, analysis of audio features such as Short-Time Fourier Transforms (STFT) and Mel Spectrograms, and utilization of several machine learning algorithms, including Least Squares Linear Regression, Random Forests (RF), Support Vector Machines (SVM), and a Convolutional Neural Network (CNN) to evaluate the usefulness of nonverbal biometrics in predicting judges’ scores. From the existing IBM Debater dataset of recorded speeches, a subset of 72 speeches across 9 speakers was selected and scored by a team of qualified school-level adjudicators to create the dataset used in these experiments. The preliminary results on the dataset were promising and have provided valuable insights into the challenges and efficacy of various regression techniques in audio-based persuasiveness prediction, highlighting the need for further exploration in this domain.
KW - Action Quality Assessment
KW - Convolutional Neural Networks
KW - Least Squares Linear Regression
KW - Random Forests
KW - Speech Processing
KW - Structured Debating
KW - Support Vector Regressor
UR - https://www.scopus.com/pages/publications/105008762420
U2 - 10.1007/978-3-031-93061-4_7
DO - 10.1007/978-3-031-93061-4_7
M3 - Conference contribution
AN - SCOPUS:105008762420
SN - 9783031930607
T3 - Lecture Notes in Computer Science
SP - 93
EP - 107
BT - Human-Centered Design, Operation and Evaluation of Mobile Communications - 6th International Conference, MOBILE 2025, Held as Part of the 27th HCI International Conference, HCII 2025, Proceedings
A2 - Wei, June
A2 - Margetis, George
PB - Springer Science and Business Media Deutschland GmbH
T2 - 6th International Conference on Design, Operation and Evaluation of Mobile Communications, MOBILE 2025, held as part of the 27th HCI International Conference, HCII 2025
Y2 - 22 June 2025 through 27 June 2025
ER -