Authors : S.D. Umarani, R.S.D. Wahidabanu and P. Raviram
Abstract: In speech recognition systems, obtaining good performance in noisy environments still remains a very challenging task. The problem is that recognition accuracy degrades significantly if training conditions are not matched to the corresponding test conditions. This study uses auditory transform along with CFCC (Cochlear Filter Cepstral Coefficients). Usually, the performance of acoustic models trained in clean speech drops significantly when tested in noisy speech. The CFCC features have shown strong robustness in this kind of situation. The auditory transform replaces the STFT in CFCC for overcoming the STFT’s disadvantage of fixed time-frequency resolution. Thus, a kind of good anti-noisy speech feature coefficient was obtained. In order to enhance the ability to resist the noises of different environments, an adaptive enhancement approach is introduced. The CFCC features with wavelet are applied to a speaker identification task to address the acoustic mismatch problem between training and testing environments. Finally, this experimental results show that noise resilience of the proposed method under small samples circumstance is better than exiting methods at least by 3 dB in worsetcase for lesser word count and at least 1 dB for larger word count. It is observed that CFCC feature with adaptive enhancement can remain better robust to noise. And its performance is more effective under low SNRs. It makes the speech recognition become possible under these conditions.
S.D. Umarani, R.S.D. Wahidabanu and P. Raviram, 2013. An Adaptively Enhanced Auditory Transform Based Feature Extraction Algorithm for Robust Speaker Identification. International Journal of Soft Computing, 8: 56-62.