Bayesian Methods for Speech Enhancement I. Andrianakis P. R. White Signal Processing and Control Group Institute of Sound and Vibration Research University of Southampton
Progress from last meeting We have gathered a number of existing Bayesian methods for speech enhancement… …added a number of our own ideas… …and compiled a framework of Bayesian algorithms with different priors and cost functions. The above algorithms were implemented and simulations were carried out to assess their performance.
Elements of Bayesian Estimation A central concept in Bayesian estimation is the posterior density LikelihoodPrior
Elements of Bayesian Estimation II Another important element is the selection of the cost function which leads in to different rules Square Error Cost Function MMSE Uniform Cost Function MAP
Motivation for this work A number of successful Bayesian algorithms already existing in the literature… Ephraim : MMSE in the Amplitude domain with Rayleigh priors Rainer : MMSE in the DFT domain with Gamma priors Lotter : MAP in the Amplitude domain with Gamma priors Some of our ideas fitted in the framework that seemed to be forming. It was interesting to “complete” the framework and test the algorithms for ourselves!
What have we examined Estimation Rules: MMSE MAP Domains: Amplitude DFT Likelihood (Noise pdf ): Gaussian
Priors - Chi Below are a number of instances for the Chi priors Strictly speaking the 2-sided Chi pdf is shown above. The 1-sided Chi is just the right half x2
Priors - Gamma …and a number of instances for the Gamma priors Note that the Gamma pdf is spikier than the Chi for the same value of
Categorisation of the examined algorithms DFTAmp Domain : Chi Prior : Chi Gamma MMSE MAP MMSE MAP Rule : In all the above algorithms can be either fixed or estimated adaptively.
Results In the following we will present results from simulations performed with the above algorithms We will first show results for fixed prior shapes. Finally, we will examine the case when the priors change shape adaptively.
Results for DFT algorithms and fixed Input SegSNR was 0 dB. Graphs for other input SNRs look similar SegSNRPESQ
SegSNRPESQ Results for AMP algorithms and fixed
Audio samples and spectrograms In the following we shall present some audio samples and spectrograms of enhanced speech with the so far examined algorithms. The clean and the noisy speech segments used in the simulations are presented below Clean SpeechNoisy Speech
Chi - DFT = 1.5 SNR = 7.17 PESQ = 2.42 SNR = 6.98 PESQ = 2.25 = 0.1 SNR = 8.61 PESQ = 2.41 SNR = 8.78 PESQ = 2.44 = 0.5 SNR = 8.62 PESQ = 2.44 SNR = 8.62 PESQ = 2.44 MMSE MAP
Gamma - DFT Gamma - DFT = 1.5 SNR = 8.65 PESQ = 2.44 SNR = 8.37 PESQ = 2.38 = 0.1 SNR = 8.85 PESQ = 2.33 SNR = 8.97 PESQ = 2.42 = 1.0 SNR = 8.24 PESQ = 2.31 SNR = 8.81 PESQ = 2.44 MMSE MAP
Chi - AMP Chi - AMP = 0.1 SNR = 9.31 PESQ = 2.41 SNR = 9.43 PESQ = 2.48 = 0.5 SNR = 8.88 PESQ = 2.47 SNR = 8.88 PESQ = 2.44 = 1.0 SNR = 8.12 PESQ = 2.35 SNR = 8.71 PESQ = 2.44 MMSE MAP
Gamma - AMP = 0.1 SNR = 9.28 PESQ = 2.34 = 0.5 SNR = 9.26 PESQ = 2.40 = 1.8 SNR = 8.99 PESQ = 2.39 MAP
Results revisited
MMSE algorithms reduce the background noise, especially for low SNRs Some examples follow… Results for adaptive MAP algorithms do not seem to improve their performance with adaptive values of
Results for adaptive = 0.05 SNR = 8.89 PESQ = 2.42 SNR = 8.96 PESQ = 2.5 = 0.3 SNR = 8.99 PESQ = 2.42 SNR = 9.07 PESQ = 2.5 = 0.1 SNR = 9.54 PESQ = 2.52 SNR = 9.43 PESQ = 2.48 Fixed Adaptive MMSE Chi Dft MMSE Gamma DftMMSE Chi Amp