Application of Perceptual Filtering Models to Noisy Speech Signals Enhancement

Table 3

Recognition rate for the proposed perceptual generalized spectral subtraction method based on the gammatone and the gammachirp filter banks decomposition (MMSE_GT and MMSE_GC) compared with the spectral modification rule (MMSE).

Babble noise
Methods/SNR	0	5	10	15
MMSE	64.81%	79.41%	89.36%	90.13%
MMSE_GT
MCRA	67.69%	84.79%	93.18%	94.97%
MCRA2	65.90%	83.74%	92.24%	93.93%
IMCRA	68.76%	84.41%	93.80%	94.62%
MS	66.24%	82.15%	92.07%	93.07%
MMSE_GC
MCRA	57.36%	74.40%	87.66%	93.00%
MCRA2	61.29%	74.73%	87.32%	92.89%
IMCRA	67.91%	84.08%	91.83%	93.24%
MS	65.78%	75.33%	90.33%	91.10%
Car noise
Methods/SNR	0	5	10	15
MMSE	75.88%	85.25%	91.40%	92.96%
MMSE_GT
MCRA	80.68%	89.56%	92.64%	95.13%
MCRA2	79.03%	86.83%	91.51%	93.84%
IMCRA	82.40%	90.51%	92.06%	94.99%
MS	79.78%	90.15%	91.88%	93.35%
MMSE_GC
MCRA	74.77%	86.71%	91.95%	93.69%
MCRA2	72.64%	82.49%	90.33%	93.08%
IMCRA	80.18%	90.20%	92.29%	94.89%
MS	79.13%	89.75%	92.57%	93.80%