Acoustic Analysis of Whispery Voice Disguise in Mandarin Chinese

Cuiling Zhang, Bin Li, Si Chen, Yike YangInterspeech 2018 · DOI

Abstract

Whispering is a common non-electronic disguise used in forensic contexts to conceal identity. This study investigates the acoustic changes that occur when speakers switch from a normal to a whispered voice in Mandarin Chinese. Analysis reveals significant shifts in vowel formants, speech rate, and intensity. Furthermore, the study evaluates the impact of these changes on human perception and automatic speaker recognition (ASR) systems, demonstrating a substantial degradation in identification accuracy. Linear regression models are proposed to compensate for these acoustic shifts, offering a potential tool for forensic voice comparison.

Introduction

In forensic science, identifying a suspect through voice recordings is a critical task. However, the use of voice disguise, particularly whispering, presents a significant hurdle. Whispering is characterized by the absence of vocal fold vibration, which eliminates the fundamental frequency ($F_0$), a primary feature for speaker identification. This study quantifies the acoustic parameters of whispered Mandarin Chinese to understand how this disguise affects both human and machine recognition.

Methodology

The study utilized a database of disguised voices from 11 male college students. Participants recorded 10 sentences typical of kidnapping scenarios in both normal and whispered voices. Acoustic analysis was performed using Praat, focusing on syllable duration, intensity, vowel formants ($F_1$–$F_4$), and Long Term Average Spectrum (LTAS). Statistical significance was assessed via paired T-tests, and linear regression was employed to model the relationship between normal and whispered parameters.

Acoustic Analysis

The transition to whispered speech involves significant physiological and acoustic changes.

Formant Shifts and Vocal Tract Narrowing

The most dramatic change was observed in vowel formants. $F_1$ increased by an average of 75%, while $F_2$ rose by 30%. These shifts are attributed to the narrowing of the vocal tract and the involvement of false vocal folds during whispering, which alters the resonance properties of the vocal tract.

Waveform and spectrogram of a normal Mandarin voice sample.

Temporal and Intensity Changes

Whispering resulted in a 14% increase in syllable duration, indicating a slower speech rate. This suggests that whispering requires greater articulatory effort or different breath management than normal phonation. Additionally, whispered speech was approximately 13 dB quieter than normal speech, representing a 22% decrease in intensity.

Spectrogram of a whispered voice sample.

Impact on Recognition Performance

The acoustic shifts caused by whispering severely impact recognition performance for both automated systems and human listeners.

  • Automatic Speaker Recognition (ASR): The ASR system failed completely when matching whispered samples to normal ones, yielding a 0% success rate. Similarity scores plummeted from near 1.0 (normal-to-normal) to well below the 0.80 forensic threshold.
  • Human Perception: Even listeners familiar with the speakers saw their recognition accuracy drop from 100% to 56% when presented with whispered speech.

Conclusion and Future Work

Whispering fundamentally alters the acoustic profile of the voice, making identification difficult. The study successfully developed linear regression models to predict normal voice parameters from whispered ones, providing a framework for parameter compensation in forensic investigations. Future research should expand the dataset to include diverse demographics and evaluate modern deep-learning-based recognition systems.

Summary updated

BibTeX
@inproceedings{zhang2018acoustic,
  author = {Cuiling Zhang and Bin Li and Si Chen and Yike Yang},
  title = {{Acoustic Analysis of Whispery Voice Disguise in Mandarin Chinese}},
  booktitle = {Interspeech},
  year = {2018},
  doi = {10.21437/interspeech.2018-2598},
}
Figure