Reproducibility Package for Controlled Robustness Engineering in Classical MFCC–LBG Speaker Identification
Description
This repository provides the reproducibility materials supporting the manuscript “Engineering Robustness in Classical Speaker Identification under Additive Noise: A Controlled Comparison of Feature-Level, Scoring, and Codebook Interventions.” The study uses classical mel-frequency cepstral coefficient (MFCC) and Linde–Buzo–Gray (LBG) vector-quantization speaker identification as a controlled platform for comparing robustness interventions introduced at different processing stages. The repository contains MATLAB source code for the baseline MFCC–LBG system, voice activity detection plus cepstral mean and variance normalization (VAD+CMVN), global diagonal Mahalanobis scoring, particle swarm optimization (PSO)-assisted LBG codebook optimization, and paired Wilcoxon signed-rank analysis with Holm adjustment. It also includes derived region-level and fold-level experimental outputs for TIMIT dialect regions DR4 and DR6, condition-preparation timing records, the statistical-comparison output, and the workbook used to derive the normalized robustness results reported in Table 4 and Figure 2. Experiments cover clean speech and additive white Gaussian noise (AWGN) and babble noise at 20, 10, and 5 dB SNR using six matched cross-validation folds. The original TIMIT speech corpus is not redistributed because access and redistribution are subject to the licensing terms of the Linguistic Data Consortium (LDC93S1). Researchers with authorized TIMIT access can use the released code and documentation to reproduce the experimental workflow. The package is intended to provide a transparent provenance chain from the experimental implementation and fold-level outputs to the statistical and normalized-robustness evidence reported in the manuscript.
Files
Steps to reproduce
1. Obtain authorized access to the TIMIT Acoustic-Phonetic Continuous Speech Corpus (LDC93S1). The corpus itself is not included in this repository. 2. Make TIMIT available locally and configure the timitRoot variable in run_timit_kfold_noise_experiment_corrected.m if the local corpus location differs from the released default. 3. Execute the main MATLAB experiment workflow for TIMIT dialect regions DR4 and DR6. The released implementation evaluates six matched cross-validation folds under clean speech, AWGN, and babble noise at 20, 10, and 5 dB SNR. 4. The experiment produces region-level summary files, fold-level result files, and separate condition-preparation timing records. 5. Run analyze_pscm_holm.m using the DR4 and DR6 fold-level CSV files to reproduce the paired Wilcoxon signed-rank comparisons and Holm-adjusted results contained in PSCM_Wilcoxon_Holm.csv. 6. Use Table4.xlsx to inspect the calculation provenance for normalized accuracy-retention and EER-inflation ratios and the degradation slopes used in Table 4 and Figure 2 of the manuscript. 7. See README.md, experimental_protocol.md, and manuscript_output_mapping.md for detailed workflow, environment, and manuscript-output provenance information.
Institutions
- Manado State UniversityNorth Sulawesi, Kelurahan Tounsaru