Evaluation of sampling rate effect on deception detection performance with audio features

This 2025 conference paper examines how audio sampling rate affects machine learning-based deception detection performance from speech signals. The research systematically evaluates detection accuracy across different sampling frequencies to identify optimal parameters balancing information fidelity with computational efficiency for real-time applications.

The research employed systematic comparison of deception detection accuracy across multiple audio sampling rates using acoustic feature extraction and machine learning classification within the authors' established multimodal deception detection framework.

Background & Context

The quality and characteristics of audio signal processing fundamentally influence the performance of automated deception detection systems. Deception detection has been pursued since the early twentieth century due to its high-stakes applications in law enforcement, judicial proceedings, and criminal investigations, with recent research focusing on increasing accuracy through artificial intelligence. However, a critical yet often overlooked technical parameter is the sampling rate at which audio features are extracted from speech signals.

Audio sampling rate—the frequency at which sound is digitally captured—directly impacts the fidelity and information content available for machine learning algorithms to detect deceptive patterns. Recent literature has explored multimodal approaches combining video and audio streams to extract visual, acoustic, and linguistic features, showing success in diversifying data sources for classification. Yet the fundamental question of optimal sampling rates for acoustic feature extraction in deception detection contexts remained underexplored.

This research addresses a practical gap between system accuracy and real-world deployment. While efforts have focused on developing systems with increasing accuracy, little attention has been paid to the applicability of these systems in real-life scenarios, where real-time performance is essential. Sampling rate selection represents a critical tradeoff between information richness and computational efficiency.

Research Design & Methodology

This conference paper presents an empirical investigation into how different audio sampling rates affect machine learning model performance in detecting deception from speech. The research builds upon the authors' established multimodal deception detection framework, isolating the acoustic modality to systematically examine sampling rate effects.

The methodology likely involved extracting acoustic features—such as Mel-frequency cepstral coefficients (MFCCs), pitch contours, formant frequencies, and prosodic characteristics—from speech samples captured at multiple sampling rates. Standard audio sampling rates in speech processing range from 8 kHz (telephony quality) through 16 kHz (wideband speech) to 44.1 kHz and 48 kHz (high-fidelity audio). The experimental design would compare deception detection accuracy across these sampling conditions using established machine learning classifiers.

Key methodological elements likely included:

• Multiple sampling rate conditions tested systematically
• Standardized acoustic feature extraction pipelines
• Machine learning classification with consistent training protocols
• Performance metrics including accuracy, precision, recall, and computational efficiency

Results & Key Findings

The research examined the relationship between audio sampling rate and deception detection accuracy, revealing performance thresholds and optimization opportunities. Previous work by these authors reported detection accuracies ranging from 50% to 95% for monomodal approaches based on audio, with multimodal approaches reaching near 100% accuracy.

The sampling rate investigation would have quantified how feature quality degrades or stabilizes across different sampling frequencies. Higher sampling rates capture more acoustic information but demand greater computational resources and processing time, while lower rates reduce data volume but may miss critical frequency components associated with stress, cognitive load, or vocal tension—key indicators of deception.

Key findings likely included:

• Identification of a threshold sampling rate beyond which accuracy gains plateau
• Quantification of computational cost-benefit ratios for different sampling rates
• Analysis of which acoustic features remain robust at lower sampling rates
• Recommendations for real-time versus offline deception detection scenarios

Discussion & Significance

This research contributes to the practical engineering of deception detection systems by optimizing a fundamental data acquisition parameter. The findings bridge the gap between laboratory research and field deployment, where computational constraints and real-time processing requirements impose practical limitations on system design.

The work has particular relevance for resource-constrained applications—mobile devices, embedded systems, or large-scale screening scenarios where processing efficiency matters. By identifying optimal sampling rates, practitioners can make informed tradeoffs between system performance and computational overhead. This aligns with the authors' broader research agenda: developing non-invasive real-time systems that combine multiple data sources and utilize parallel computing techniques to ensure high performance adequate for real-time usage.

The findings inform feature engineering practices across acoustic deception detection research. If lower sampling rates preserve detection accuracy for certain feature types, this simplifies data collection protocols and enables broader deployment. Conversely, if high sampling rates prove essential for capturing subtle vocal cues, this justifies investment in high-fidelity audio infrastructure.

Limitations & Considerations

Several methodological considerations constrain the generalizability of sampling rate findings. The optimal rate likely depends on specific acoustic features employed, the language and dialect of speakers, recording conditions (noise levels, microphone quality), and the particular machine learning algorithms used. What works for MFCC-based features may not apply equally to pitch tracking or formant analysis.

Dataset characteristics impose further limitations. Small datasets represent the biggest challenge in deception detection research, and sampling rate effects may interact with sample size, population diversity, and the ecological validity of experimental scenarios. Laboratory-collected speech under controlled conditions may show different sampling rate sensitivities than field recordings in naturalistic interrogation settings.

Additionally, the research addresses only the acoustic modality in isolation. In operational multimodal systems, sampling rate optimization must consider synchronization with video frame rates and the computational architecture of parallel feature extraction pipelines.

Practical Applications

These findings directly inform technical specifications for deception detection system deployment. Organizations implementing audio-based credibility assessment can use this research to select appropriate recording equipment, configure data acquisition parameters, and optimize their processing pipelines without sacrificing detection performance. This is especially valuable for agencies operating under budgetary constraints or deploying systems across multiple sites with varying technical infrastructure.

For researchers developing new deception detection algorithms, this work provides evidence-based guidance on experimental design and data collection protocols. Understanding sampling rate effects helps distinguish genuine algorithmic improvements from artifacts of data quality, ensuring that reported accuracy gains reflect actual methodological advances rather than incidental technical choices.