A forensically-motivated tool for selecting cepstrally-consistent steady-states from non-contemporaneous vowel utterances

Michael Barlow, Mehrdad Khodai-Joopari, Frantz ClermontInterspeech 2004

TL;DR

This paper introduces a semi-automatic tool designed to extract consistent and robust "steady-state" vowel segments from speech recordings, even when those recordings are from different times. This is crucial for forensic speaker identification, where subtle acoustic cues need to be reliably identified across potentially varied recording conditions.

What Problem Does This Paper Address?

Forensic speaker identification (FSI) relies on analyzing unique acoustic characteristics of a speaker's voice. A key challenge in FSI is ensuring that the speech features used for comparison are consistent and reliable, especially when comparing recordings made at different times (non-contemporaneous). Vowel sounds, particularly their "steady-state" portions (where the vocal tract is relatively stable), are known to contain significant speaker-specific information. However, automatically and reliably identifying these steady-state segments, especially across recordings with varying noise or recording quality, is difficult. Previous methods often relied on simple spectral variation, which could be sensitive to non-contemporaneity and noise, leading to inconsistent results. This paper aims to address this by developing a more robust method for selecting these crucial vowel segments.

Methods

The authors propose a two-phase tool:

  • Phase 1: Interactive and Semi-Automatic Procedures: This phase focuses on identifying a continuous, high-energy interval within a vowel utterance that is relatively free of noise and transients. It uses an energy curve to initially highlight potential regions and then employs a "Within Token Frame Cepstral Distance" (WTFCD) measure. WTFCD leverages the "peak-sensitivity property" of an index-weighted cepstral distance, which is less sensitive to channel effects and more attuned to actual spectral peak movements. This helps in identifying potential steady-state candidates within the selected interval. The process is interactive, allowing users to refine the selection.
  • Phase 2: Fully Automatic Procedure: This phase addresses the challenge of non-contemporaneity and within-speaker variability. It takes the candidates identified in Phase 1 and uses a "Between Token Cepstral Similarity" (BT_CEP_SIM) measure. The goal is to find combinations of steady-state candidates from different recordings of the same speaker that exhibit the highest spectral similarity. This is achieved by constructing all possible combinations of candidates (four per token, four tokens per speaker) and selecting the combination with the global minimum in BT_CEP_SIM, aiming for spectral homogeneity across non-contemporaneous recordings.

Datasets

The speech materials were sourced from the "Speaker Database of the Japanese Research Institute for Police Science" (NRIPS). This database contains two non-contemporaneous recordings (separated by 3-4 months) of 300 adult male native Japanese speakers. Each recording includes two repetitions of the five Japanese vowels pronounced in isolation, along with words and sentences. The recordings were sampled at 10 KHz and quantized to 12 bits.

Key Insights and Ideas

  • Cepstral Consistency for Non-Contemporaneity: The core idea is that by focusing on cepstral similarity, especially using a peak-sensitive distance measure, the tool can find steady-state segments that are acoustically similar even if recorded at different times.
  • Peak Sensitivity for Robustness: The use of an index-weighted cepstral distance that emphasizes spectral peak differences makes the method more robust to variations in the transmission channel or recording equipment.
  • Two-Phase Approach for Refinement: The separation into an interactive phase for initial selection and a fully automatic phase for cross-recording consistency allows for both user control and automated robustness.
  • Vowel Steady-States as Rich Information: The paper reinforces the idea that steady-state vowel segments, less affected by coarticulation, contain significant speaker-specific information, making them valuable for FSI.

Key Results

The paper presents a tool that successfully identifies steady-state vowel segments. While specific quantitative results (e.g., identification accuracy improvements) are not detailed in this excerpt, the paper illustrates the effectiveness of Phase 2 by showing cepstrally-smoothed NDPS graphs for min-based and max-based variance frames. These graphs visually demonstrate that the selected candidates with minimum variance exhibit lower cepstral variability across the frequency spectrum compared to those with maximum variance, indicating successful achievement of spectral homogeneity. The authors report that their tool has yielded a moderately large dataset of steady-state segments for 60 speakers so far.

Limitations and Future Work

  • Limitations: The paper notes that the number of candidates selected (four) was determined through trial and error and might need adjustment for signals of different durations. A lesser number of candidates could mean fewer combinations but also a higher chance of selecting frames further from the true representative pattern.
  • Future Work: The authors plan to complete the parameterization for the remaining 240 speakers in the NRIPS database. They also intend to further study cepstral variability and its implications for forensic speaker identification, suggesting ongoing research into the practical application of their tool.

Summary updated

BibTeX
@inproceedings{barlow2004forensically,
  author = {Michael Barlow and Mehrdad Khodai-Joopari and Frantz Clermont},
  title = {{A forensically-motivated tool for selecting cepstrally-consistent steady-states from non-contemporaneous vowel utterances}},
  booktitle = {Interspeech},
  year = {2004},
}

Loading related content...

Figure