Data selection in forensic automatic speaker comparison

Area of Science:

• Forensic Science
• Speech Technology
• Acoustics

Background:

• Automatic speaker recognition (ASR) is crucial for forensic speaker comparison (FSC) casework.
• ASR software generates a case score, requiring calibration to a likelihood ratio (LR) for evidence strength.
• Calibration requires a representative audio dataset (LR calculation set), but selection criteria are subjective.

Purpose of the Study:

• To investigate the effect of various audio conditions on ASR score-to-LR calibration.
• To provide data-driven guidance for forensic practitioners selecting calibration sets.
• To analyze the influence of the number of speakers in the calibration set on variability.

Main Methods:

• Utilized forensically relevant audio data.
• Investigated multiple audio conditions including duration, ethnolect, recording environment (car), and data type (operational vs. lab).
• Analyzed the impact of speaker count within the LR calculation set.

Main Results:

• Duration differences below 40 seconds significantly impact score-to-LR mappings.
• Ethnolect differences, car recordings, and speech durations over 40 seconds showed no significant impact.
• Gender and operational/lab data differences were not impactful, except during cross-comparisons.

Conclusions:

• Audio duration is a critical factor in ASR calibration for FSC.
• Forensic practitioners should account for duration differences under 40s when selecting calibration sets.
• Other investigated conditions (ethnolect, car, longer durations, gender, data type) appear less critical, simplifying calibration set selection.