Data strategies in forensic automatic speaker comparison

Area of Science:

• Forensic Science
• Speech Technology
• Biometrics

Background:

• Automatic speaker recognition (ASR) is crucial for forensic speaker comparison (FSC).
• ASR requires representative audio data for reference normalization and score-to-LR function training.
• Stable ASR performance necessitates a minimum number of speakers, traditionally 30 for calibration and 60 for both calibration and normalization.

Purpose of the Study:

• To investigate data strategies for ASR in FSC when only 30 speakers are available.
• To evaluate methods for overcoming data limitations in forensic speaker recognition.
• To determine if reduced data requirements can maintain or improve ASR performance in FSC.

Main Methods:

• Simulated a scenario with only 30 available speakers.
• Tested data strategies: omitting reference normalization, splitting speakers into smaller groups, and a leave-1-or-2-out approach.
• Compared these strategies against a baseline using the required 60 speakers.

Main Results:

• The leave-1-or-2-out strategy with 30 speakers performed comparably to the baseline (60 speakers).
• Extending the leave-1-or-2-out strategy to 60 speakers resulted in performance exceeding the baseline.
• The proposed strategies significantly reduce the data requirements for ASR in FSC.

Conclusions:

• A viable data strategy can halve the data needs for ASR in forensic speaker comparison.
• This approach makes ASR more feasible in FSC casework with limited speaker data.
• Reduced data requirements enhance the applicability of ASR in forensic contexts.