Abstract: Deception detection in speech is a critical area of research with applications in security and law enforcement. While English-language corpora are relatively common, Mandarin Chinese presents unique challenges due to its tonal nature and cultural differences. This paper details the construction of the SUSP-DSD (Soochow University Speech Processing - Deceptive Speech Detection) corpus, a specialized dataset designed to bridge this gap. We describe the three-part collection paradigm and provide both subjective and objective evaluations of the resulting data.
Introduction
Traditional deception detection methods, such as polygraphs, are often invasive and can induce stress-related artifacts in subjects. Speech-based detection offers a non-invasive alternative. However, the lack of high-quality, culturally relevant datasets for non-English languages—specifically Mandarin—has hindered progress. Mandarin's use of pitch to distinguish lexical meaning makes the acoustic analysis of deception particularly complex. The SUSP-DSD corpus addresses this by providing a controlled yet ecologically valid dataset of truthful and deceptive Mandarin speech.
Methodology
The corpus was developed using 40 native Mandarin speakers (20 male, 20 female) from Soochow University. To capture a range of deceptive behaviors, a three-part "Collection Paradigm" was employed:
- Controlled Binary Responses: Participants answered "shi" (is) or "bushi" (is not) to basic factual questions, providing direct comparisons of identical words in truthful and deceptive contexts.
- Phonetic Vowel Coverage: Subjects identified numbers (1–10) on cards. These numbers were selected to ensure coverage of core Mandarin vowels ('a', 'e', 'i', 'u'), facilitating the study of deception's impact on specific phonemes.
- High-Stakes Storytelling: Participants prepared personal narratives (one true, one false). To induce realistic cognitive load and emotional arousal, a 100-yuan incentive was offered for successfully deceiving interviewers.
The SUSP-DSD Corpus
The resulting corpus contains high-quality recordings (originally 48kHz, downsampled to 16kHz) meticulously labeled for research use. Each file follows a specific naming convention to denote the test part, participant ID, and truth/deception status.
Analysis and Results
Evaluation of the corpus involved both human subjective testing and objective acoustic analysis.
Subjective Evaluation
Twelve student evaluators attempted to distinguish lies from truths in the storytelling segment. The average human accuracy was 56.86%, only slightly better than chance. This low accuracy suggests that the deceptive speech in the corpus is of high quality and lacks obvious "tells" that would make it trivial to detect, thereby providing a challenging benchmark for automated systems.
Objective Acoustic Analysis
Researchers focused on pitch-related features, including maximum pitch ($P_{max}$), mean pitch ($P_{mean}$), pitch change rate ($P_{rate}$), and standard deviation ($P_{std}$). Analysis revealed that deceptive speech often exhibits higher emotional arousal and cognitive effort, manifested as increased pitch variance and visible hesitation in pitch tracks.
Quantitative comparisons showed significant differences. For instance, a deceptive utterance from one participant showed a $P_{max}$ of 158.93 Hz compared to 120.39 Hz for a truthful utterance, alongside a higher standard deviation (17.77 vs. 12.93).
Conclusion
The SUSP-DSD corpus provides a foundational resource for Mandarin deceptive speech detection. By combining controlled phonetic tasks with high-stakes storytelling, the dataset captures the nuanced acoustic markers of deception in a tonal language. Future work will involve scaling the participant pool and applying advanced machine learning classifiers to establish automated baselines.
Summary updated