Band-Limited Cepstral Analysis of Speaker Sensitivity in Forensic Voice Comparison

TL;DR — This paper investigates the distribution of speaker-discriminative information across different frequency sub-bands using Band-Limited Cepstral Coefficients (BLCCs) for forensic voice comparison, finding that the 0–0.6 kHz and 4–6 kHz regions provide the strongest speaker cues while frequencies above 7 kHz are the least informative.

Key contributions

• Evaluated the speaker-discriminative information of individual sub-bands across the 0–8 kHz full-band spectrum using systematic 0.6 kHz and 1.2 kHz BLCC scans.
• Demonstrated that while all sub-bands yield useful speaker-specific information (Cllr < 1), speaker sensitivity is highly non-uniform, peaking in the 0–0.6 kHz and 4–6 kHz ranges.
• Showed that fusing likelihood ratios from strategically selected, non-overlapping sub-bands significantly improves forensic voice comparison performance compared to individual sub-bands.
• Established a flexible linear transformation pipeline for extracting sub-band cepstral coefficients without needing to re-analyze the original acoustic signals.

Problem

Prior forensic voice comparison studies disagree on precisely which frequency bands carry the most speaker-specific information, with existing works pointing variously to low, mid, or mid-to-high frequency regions. Furthermore, conventional sub-band analyses require computationally expensive re-processing of the original speech signal for every frequency range of interest. Understanding these spectral distributions is crucial for providing transparent, interpretable evidence in forensic contexts.

Method

The study employs Band-Limited Cepstral Coefficients (BLCCs) derived via a three-step pipeline: short-time filter-bank spectral analysis with a linear frequency scale (25 triangular filters, 25-ms frames, 5-ms shift), extraction of 14 full-band linear-frequency cepstral coefficients (LFCCs), and a linear transformation matrix converting full-band cepstra to sub-band cepstra. The 0–8 kHz frequency range was systematically scanned using 0.6 kHz sub-bands (shifted by 0.2 kHz, yielding 2 BLCC coefficients) and 1.2 kHz sub-bands (shifted by 0.4 kHz, yielding 3 BLCC coefficients).

Experiments were conducted within a Gaussian Mixture Model–Universal Background Model (GMM-UBM) framework utilizing diagonal-covariance models with 256 Gaussian components and mean-only MAP adaptation. System scores were converted to likelihood ratios using logistic-regression calibration across a 6-fold rotation scheme.

Experimental setup

Evaluated on pseudo-police interview recordings from the AusEng 500+ database comprising 30 seconds of running speech per session from 162 native male Australian English speakers. Systems were compared against a full-band LFCC baseline (Cllr = 0.13643, Ccal_llr = 0.02894, EER = 3.396%). Evaluation metrics include the log-likelihood-ratio cost (Cllr), its calibration component (Ccal_llr), and equal error rate (EER).

Results

All individual sub-bands achieved Cllr values below 1.0, but performance varied sharply across the spectrum. For 0.6 kHz sub-bands, the lowest Cllr (strongest discrimination) occurred at the 0–0.6 kHz band (0.58339) and a secondary prominent dip around 4–5 kHz, whereas frequencies above 7 kHz showed the weakest performance, with the final sub-band reaching Cllr values up to 0.82571. Fusing the likelihood ratios of up to four well-spaced, highly informative sub-bands substantially outperformed any individual sub-band and improved upon the full-band LFCC baseline.

Limitations

The study is restricted to 162 native male speakers of a single accent (Australian English) using clean pseudo-interview speech sampled at 16 kHz, limiting generalizability to female speakers, cross-linguistic scenarios, and telephony channels. Sub-band combinations were constrained to a maximum of four sub-bands due to computational limitations in the fusion analysis.

Why read this

Speech and forensic audio researchers will find this paper valuable for its rigorous mapping of speaker-discriminative cues across frequency bands using BLCCs, offering a blueprint for building multi-band or weighted feature-fusion systems.

Code

None released (as of this page's updated date). If you are an author with a repo, please claim this entry — see CONTRIBUTING.md.

Applications

Forensic voice comparison, speaker verification, and multi-band acoustic feature weighting for speaker recognition.

Institutions

Australian National University, Daido University

Related

• Sub-band Cepstral Analysis of Speaker-Specific Information: A Case Study of Japanese Word /saN/ — shared technique · relatedness 3.0/3
• Function words: a topic independent approach to word n-gram selection for forensic speaker comparison — same problem · relatedness 1.7/3
• Codec-induced Mismatch, Speech Duration, and Speaker-dependent Effect in a DNN-based Forensic Speaker Recognition System — same problem · relatedness 1.6/3
• Human-like cross-language generalisation in deep neural speaker embeddings and its acoustic foundations — same problem · relatedness 1.6/3
• Do speech foundation models perceive speaker similarity as humans do? — same problem · relatedness 1.6/3

All 950k paper pairs scored by TypeSafe Jev (scripts/related/); relatedness 0 = unrelated … 3 = directly comparable.