Elamur Research⌕
☰
home / papers / 2021 / A Study of Acoustic Features in Arabic Speaker Identification under Noisy Environmental Conditions

A Study of Acoustic Features in Arabic Speaker Identification under Noisy Environmental Conditions

Summary

The paper asks which acoustic front-end keeps a speaker identification system reliable when the test speech is degraded by environmental noise, focusing on Arabic speakers. Five feature types are compared: MFCC, GFCC, PNCC, PLP and LSF. A standard GMM back-end with 16 mixtures is used, and the authors also test pairwise fusions of the stronger features. Experiments use the ARADIGIT corpus of Arabic spoken digits from Algerian speakers. Speaker models are trained on clean speech. Test utterances are artificially corrupted with babble, factory and subway noise from NOISEX'92 at SNRs from −6 to 18 dB.

The two auditory-inspired features, GFCC and PNCC, clearly beat the conventional MFCC baseline in all three noise types, and GFCC is the best overall. Both rely on a gammatone filter bank. PLP ranks third. Combining the strong features did not add robustness. Subway noise hurts identification the most. The study is a practical guide to choosing features for noisy biometric speech systems in a less-studied language.

Key points

  • Data: ARADIGIT contains 62 speakers (31 male, 31 female), recorded at 22.05 kHz and downsampled to 16 kHz.
    • Each speaker repeats the digits 0–9 three times.
    • Concatenated digits 0–7 are used for training and digits 8–9 for testing.
  • Noise setup: noise was added with the FaNT tool, using NOISEX'92 babble, factory and subway noise. SNR ranged from −6 to 18 dB in 6 dB steps.
  • Feature setup:
    • PLP, GFCC and PNCC each use 13 coefficients plus deltas and delta-deltas, giving 39 dimensions.
    • MFCC uses 12 coefficients plus energy, with the same derivatives.
    • LSF uses 10 parameters.
    • The fusions (GFCC+PNCC, PLP+GFCC, PLP+PNCC) are 26-dimensional.
  • Ranking: GFCC was best, followed by PNCC and then PLP. MFCC and LSF were weakest, though MFCC beat LSF under subway noise.
  • Thresholds and fusion: PLP becomes competitive only from about 12 dB in babble and factory noise. Fused features did not outperform the single best features.

Interesting ideas

  • I1. Graded degradation sweep. Clean models are tested on inputs corrupted at fixed, evenly spaced severity levels and with several noise types, which yields a robustness curve rather than a single score. — For Mova (learn): Inject graded text noise (typos, lost punctuation, casing damage) into held-out treebank sentences at set levels. Plot how LAS degrades per noise type to locate where the parser breaks.
  • I2. Report results per condition. Each noise type harmed accuracy differently, with subway noise the worst, so pooled averages would hide the real weak spot. — For Mova (answer): Break reader accuracy down by text source (school tests, word problems, fables). This shows which genre undermines comprehension instead of relying on one overall number.
  • I3. Fusion must earn its place. Joining the two strongest feature sets did not help, so combining good components should be checked empirically rather than assumed. — For Mova (learn): Before adding another witness parser to silver-data agreement, measure whether it actually raises agreed-label accuracy. Keep only witnesses that add value.
  • I4. Estimate and remove the background floor. PNCC tracks a slowly varying noise floor over a medium time window and normalizes the signal against it, rather than treating every frame alike. — For Mova (frames): Estimate a document-level baseline of frequent, low-information constructions (stock phrases, boilerplate). Let implicature or irony detection react only to deviations from that baseline.
  • I5. Compressive nonlinearity instead of log. GFCC uses a cubic root and PNCC a 1/15 power law, both milder on near-zero values than a logarithm, and this aids robustness. — For Mova (tree): When turning raw counts or scores into features for the perceptron parser, try a power-law compression instead of log. This could reduce the instability caused by rare, low-count events.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2110.12304 · arXiv 2110.12304

@misc{benhafid2021study,
  title = {{A Study of Acoustic Features in Arabic Speaker Identification under Noisy Environmental Conditions}},
  author = {Benhafid, Zhor and Zergat, Kawthar Yasmine and Amrouche, Abderrahmane},
  year = {2021},
  doi = {10.48550/arxiv.2110.12304},
  eprint = {2110.12304},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2110.12304}
}

References

18 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A. Harrag, “Database features extraction : speaker features application (extraction des données d’une base: Application à l’extraction des traits du locuteur),” Ph.D. dissertation, Faculty of Technology, UFAS, Algeria, Jun. 2011. find
  2. J. Ming, T. Hazen, J. Glass, and D. Reynolds, “Robust speaker recognition in noisy conditions,” IEEE Trans. Audio, Speech, Lang. Process., vol. 15, no. 5, p. 1711 – 1723, 2007. find
  3. T. May, S. van de Par, and A. Kohlrausch, “Noise-robust speaker recognition combining missing data techniques and universal background modeling,” IEEE Trans. Audio, Speech, Lang. Process., vol. 20, no. 1, pp. 108–121, 2012. find
  4. G. Hern’andez, J. Calvo, R. F. andez, I. Rodes, and R. Mart’ınez, “Noise robust voice detector for speaker recognition,” in IEEE Int. Conf. On Pattern Recognition, Tampa, FL, USA, Dec. 2008. find
  5. X. Zhao, Y. Shao, and D. Wang, “Casa-based robust speaker identification,” IEEE Trans. Audio, Speech, Lang. Process., vol. 20, no. 5, p. 1608 – 1616, 2012. find
  6. X. Zhao and D. Wang, “Analyzing noise robustness of mfcc and gfcc features in speaker identification,” in IEEE Int. Conf. On Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 2013. find
  7. X. Zhao, Y. Wang, and D. Wang, “Robust speaker identification in noisy and reverberant conditions,” in IEEE Int. Conf. On Acoustics, Speech and Signal processing, Florence, Italy, 2014. find
  8. E. Tazi, “A robust speaker identification system based on the combination of gfcc and mfcc methods,” in in IEEE Int. Conf. On Multimedia Computing and Systems, Marrakech, Morocco, 2016. find
  9. S. Shafee and B. Anuradha, “Speaker identification and spoken word recognition in noisy background using artificial neural networks,” in IEEE Int. Conf. On Electrical, Electronics, and Optimization Techniques, Chennai, India, 2016. find
  10. X. Shi, H.Yang, and P. Zhou, “Robust speaker recognition based on improved gfcc,” in IEEE Int. Conf. on Computer and Communications, Chengdu, China, 2016. find
  11. C. Kim and R. Stern, “Power-normalized cepstral coefficients (pncc) for robust speech recognition,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 24, no. 7, pp. 1315 – 1329, 2016. find
  12. E. Ambikairajah, V. S. J. Min Karen Kua, and H. Li, “Pncc-ivector-src based speaker verification,” in Signal and Information Processing Association Annual Summit and Conference, Hollywood, CA, USA, 2012. find
  13. X. Jing, B. Xiang, H. Yang, and P. Zhou, “Robust speaker verification using improved pncc based on gmm-ubm,” International Journal of Automation and Power Engineering., vol. 4, 2015. find
  14. N. Dehak, “Discriminative, and generative approaches for long-and short-term speaker characteristicsmodeling: Application to speaker verificaton,” Ph.D. dissertation, Faculty of Technology, Monteral, Quebec, Jun. 2009. find
  15. H. Hermanskay, “Predictive (plp) analysis of speech,” The Journal of the Acoustical Society of America., vol. 87, no. 4, 1990. find
  16. K. Zergat and A. Amrouche, “Influence of g.729 coded bitstream in automatic speaker recognition using gaussian mixture model,” Journal *Fig. 4. SID accuracy (%) comparisons of the proposed combined system.* of Electrical and Electronics Engineering., vol. 6, no. 1, pp. 159–162, May 2013. find
  17. K. Zergat, A. Amrouche, M. Fedila, and M. Debyeche, “Robust arabic speaker verification system using lsf extracted from the g.729 bitstream,” in IEEE Int. Workshop on Machine Learning for Signal Processing, 2013. find
  18. H. G. Hirsch, “Fant-filtering and noise adding tool,” Hochschule Nider-rhein Universty of Applied Sciences, Tech. Rep., Mar. 2005. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/benhafid-2021-study-acoustic-features-arabic-speaker/