Demos of the HMM-based Analysis/Resynthesis approach to speech enhancement. For full details see,
Carmona, J.L., Barker, J., Gomez, A.,M. and Ma, N. (2013) Speech Spectral Envelope Enhancement by HMM-based Analysis/Resynthesis, IEEE Signal Processing Letters, 20(3):563-566
In the table below the abbreviations have the following meaning (see paper for details)
MI = mean imputation
CI = constrained imputation
G0 = no grammar or dictionary; G1 = using dictionary and sentence grammar
SD = speaker dependent models
Enhancement
male
female
None
Spectral Subtraction
MMSE
MI+G0
CI+G0
CI+G0+SD
CI+G1+SD
Clean
The technique is currently only enhancing the spectral envelope so there is a roughness in the quality of the enhancement that is arising from noise in the excitation. If the excitation of the original speech signal is used then this disappears as demonstrated below. A small amount of musical noise be caused by errors in the mask estimation. Note this is not present when using the oracle mask.