Validation of an embedding-based forensic voice comparison system using short speech samples in Chinese languages

Area of Science:

• Speech processing
• Forensic science
• Machine learning for biometrics

Background:

• Speaker recognition systems are crucial in forensic voice comparison.
• ECAPA-TDNN x-vector systems offer advanced speaker embedding extraction.
• Understanding system performance under varying conditions is vital for practical application.

Purpose of the Study:

• To evaluate the performance of an ECAPA-TDNN x-vector system in forensic voice comparison.
• To investigate the impact of speech duration and calibration set size on system accuracy.
• To compare performance across different languages (Hong Kong Cantonese and Northeast Mandarin).

Main Methods:

• Utilized forensically realistic speech data from 90 Hong Kong Cantonese and 90 Northeast Mandarin speakers.
• Extracted speaker embeddings using a pre-trained ECAPA-TDNN model.
• Applied Probabilistic Linear Discriminant Analysis (PLDA) scoring and logistic regression calibration.
• Systematically varied speech duration (2-20s) and calibration group size (20-60 speakers).

Main Results:

• Performance improved significantly with speech duration up to 10 seconds, with diminishing gains thereafter.
• Hong Kong Cantonese (HKC) consistently outperformed Northeast Mandarin (NEM).
• Calibration size had a minor effect compared to duration, with more fluctuation observed with fewer than 30 speakers.

Conclusions:

• Speech duration is a critical factor for ECAPA-TDNN x-vector system performance in forensic voice comparison.
• The system demonstrates robust performance, particularly with sufficient speech duration.
• Language-specific characteristics, like content similarity, can influence comparative accuracy.