Sub-band Cepstral Analysis of Speaker-Specific Information: A Case Study of Japanese Word /saN/

TL;DR — This study investigates how speaker-discriminative information is distributed across frequency sub-bands for Japanese segments /s/, /a/, and /N/ using Band-Limited Cepstral Coefficients (BLCCs) in a forensic voice comparison (FVC) framework. The vowel /a/ and nasal /N/ outperformed the fricative /s/, and fusing just four optimal sub-bands approached full-band FVC performance.

Key contributions

• Applied the band-limited cepstral coefficient (BLCC) method via linear transformations to investigate sub-band speaker-specific information for consonants (/s/, /N/) and vowels (/a/) without signal re-analysis.
• Demonstrated that speaker-discriminative information is unevenly distributed across the 0-8 kHz spectrum, with performance degrading sharply in the final 1 kHz (7-8 kHz) across all segments.
• Identified optimal sub-band locations for each phoneme: 1-2 kHz for /s/, 5-6 kHz and formant regions F2/F3 for /a/, and <4 kHz (peaking near 0.5-1 kHz) for /N/.
• Showed that fusing a small number (up to 4) of well-spaced, highly informative sub-bands yields log-likelihood-ratio costs approaching those of full-band systems.

Problem

Speaker-discriminative information is non-uniform across the audio spectrum, yet most prior forensic voice comparison (FVC) and running speech studies provide only broad, whole-spectrum views without detailing how speaker-specific cues localize across different phonemic contexts. Traditional filterbank-based sub-band analyses lack precise control and flexibility, while prior work on isolated segments has focused almost exclusively on vowels rather than consonants like fricatives (/s/) and nasals (/N/). Understanding these fine-grained spectral distributions is legally and forensically crucial for interpreting voice comparison evidence accurately in court.

Method

The authors utilized the Band-Limited Cepstral Coefficient (BLCC) method to extract sub-band cepstral representations directly from full-band features via linear transformation matrices, avoiding signal re-analysis. Speech data was analyzed using 25-ms frames with 5-ms shifts, applying 25 triangular filters on a linear frequency scale from 0 to 8 kHz to extract 14 linear-frequency cepstral coefficients (LFCCs). The three central frames were averaged to form a stable full-band vector per segment.

Sub-band analysis was performed using 0.6-kHz wide sub-bands shifted in 0.2-kHz increments across the 0–8 kHz range, yielding 38 overlapping sub-bands. The transformation parameter was set such that each BLCC vector contained two coefficients (including the 0th-order term). Likelihood ratios (LRs) were estimated using a multivariate kernel density model, calibrated via logistic regression, and evaluated using log-likelihood-ratio costs (C_llr). Experiment 2 systematically fused LRs from groups of two to four sub-bands to measure combinatorial performance gains against full-band baselines.

Experimental setup

Evaluated using the National Research Institute of Police Science database of Japanese adult-male speakers (306 speakers, mean age 39.9 years) producing citation-form tokens of the word /saN/ across two recording sessions separated by 3–5 months. A six-fold cross-validation rotation scheme was used (102 speakers per batch, split into test, background, and calibration sets), yielding 102 same-speaker and 10,302 different-speaker comparisons per fold. Performance was measured via log-likelihood-ratio cost (C_llr, C_min, C_cal) and Equal Error Rate (EER), comparing sub-band models and multi-sub-band fusions against full-band LFCC baselines.

Results

Full-band baseline performance showed that the vowel /a/ yielded the best results (C_llr = 0.38766, EER = 9.66%), followed by the nasal /N/ (C_llr = 0.43473, EER = 12.09%), while the fricative /s/ performed considerably worse (C_llr = 0.68110, EER = 22.03%). In sub-band analyses, all segments exhibited sharp performance degradation in the highest 7–8 kHz frequency band. For /s/, the best performance concentrated around 1–2 kHz and 6.5 kHz; for /a/, the peak single sub-band occurred at 5.2–5.9 kHz (C_llr = 0.76044) alongside F2/F3 regions; for /N/, performance concentrated below 4 kHz with a minimum around 0.5–1 kHz. Fusing four carefully selected sub-bands yielded C_llr values that closely approached the full-band system.

Limitations

The study is restricted to a single word token (/saN/), adult male speakers of a single language (Japanese), and laboratory-quality recordings downsampled to 16 kHz, which does not reflect severe telephone channel bandwidth constraints or channel mismatch conditions typical of real forensic casework. Furthermore, combinations exceeding four fused sub-bands were omitted due to computational constraints, and alternative filter scales (such as Mel or Bark) or linear prediction coefficients were not evaluated.

Why read this

Speech and forensic science researchers will appreciate how the BLCC framework enables precise, computationally efficient sub-band extraction without signal re-analysis. It provides clear empirical evidence that speaker-discriminative cues are highly non-uniform and phoneme-dependent.

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 phonetic profiling of speech segments.

Institutions

Australian National University

Related

• Band-Limited Cepstral Analysis of Speaker Sensitivity in Forensic Voice Comparison — shared technique · relatedness 3.0/3
• LaS-LCA: Layer-Selected Latent Cross-Attention Adapters and Margin-Mixup for Robust Cross-Lingual Speaker Verification — same problem · relatedness 1.8/3
• Beyond Short Segments : Expanding Speaker Embeddings with Vector Archives — same problem · relatedness 1.8/3
• Adapting Audio Large Language Models for Speaker Verification — same problem · relatedness 1.8/3
• Progressive Learning for Robust Speaker Representation — same problem · relatedness 1.8/3

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