A smoothed Minimum Mean-Square Error Log-Spectral Amplitude estimator for speech enhancement

Typical single-channel speech enhancement algorithms employ a multiplicative gain filter in the frequency-domain. This spectral gain is multiplied with the short-term spectrum of noisy speech in order to produce an estimate of the clean speech spectrum. More sophisticated techniques derived the calculation of this gain under various optimization criteria. This paper analyzes the use of smoothed gain filter for speech enhancement with the aim of gaining a better improvement in their performance. The smoothing technique is applied to the well-known Minimum Mean-Square Error short-time Log-Spectral Amplitude (MMSE-LSA) estimator. The performances of this technique are analyzed and compared with those of the classical MMSE-LSA estimator. Objective and subjective evaluation of the speech enhancement confirm superiority in noise reduction and quality of the enhanced speech.