Enhancement of Speech in Noisy Conditions Project Presentation Paul Coffey Contents Project Overview System Development Testing Conclusion Paul Coffey Project Presentation Project Overview This project is about The “Enhancement of Speech in Noisy Conditions” Speech Enhancement is used in communications everywhere – From Mobile Phones to Speech Recognition Systems Paul Coffey Project Presentation Project Overview To Investigate Spectral Subtraction – – A method to improve the quality of a signal that has been effected by noise This is done by taking noise from degraded signal Noisy Signal Analysis Window Fourier Transform Spectral Modification Inverse Fourier Transform Synthesis Window X ( f ) S ( f ) P( f ) Overlap Add Enhanced Signal Paul Coffey Project Presentation Project Overview Implement Wiener Filter Method – Noisy Signal Is derived and Works on same basis as Spectral Subtraction Speech and noise power spectrum estimation PXX(f) PNN(f) Wiener Filtering Pxx f W f Pxx f Pyy f Paul Coffey Wf Pxx f Pxx f Pyy f Enhanced Signal Project Presentation System Development Paul Coffey 4 4 Original Speech Sample x 10 2 Mag To begin development, a simple task such as doubling the magnitude was implemented Also getting the system to analyse and synthesise the whole signal Hamming Windows and Overlap Add were also introduced to the system 0 -2 -4 0 500 1000 1500 500 1000 1500 5 1 x 10 2000 2500 Samples New Speech signal 3000 3500 4000 3000 3500 4000 0.5 Mag 0 -0.5 -1 0 2000 Samples 2500 Project Presentation System Development Finally a basic Spectral Subtraction implementation of the filter was then achieved Result from running the filter is shown MagY = MagX – MagNN; Original , Noisy and Filtered speech 1 0.5 0 Mag -0.5 -1 -1.5 0 0.5 1 1.5 2 Time Paul Coffey Project Presentation 2.5 4 x 10 System Development The other filter that was developed was the Wiener Filter This was approached in the same way as the previous filter Since this method is similar to the way it is derived as Spectral Subtraction, this greatly helped with speeding up development Paul Coffey Project Presentation System Development Basic Wiener Filter Implementation Result of the signal being filtered shown This is implemented using the code below: W = MagX./(MagX+MagNN); MagY = MagX .* W; Paul Coffey Original, Noisy and Filtered Speech Signal 1 0.5 Mag 0 -0.5 -1 -1.5 0 0.5 1 1.5 Time 2 2.5 Project Presentation 3 4 x 10 System Development Noise Detection and SNR were also needed to be developed These were required for the filters to operate properly Finding the energy levels of the signal was used to implement the noise detection method The code to do so is: E = sum (x.^2); The SNR is then was developed using the equation SNR db 10 log Paul Coffey 10 2 x 2 n Project Presentation Testing Throughout the project tests were being carried out to see how the filters were working Spectral Subtraction here with SNR of 10db Original, Noisy and Filtered with SNR of 10db 0.8 0.6 0.4 0.2 0 Mag -0.2 -0.4 -0.6 -0.8 -1 -1.2 0 0.5 1 1.5 Samples Paul Coffey 2 2.5 4 x 10 Project Presentation Testing Here is the Wiener Filter Implementation with SNR at 10db It is the same speech signal as used in the Spectral Subtraction implementation Can see Wiener doesn’t cut out speech like the Spectral did Original, Noisy and Filtered signal with SNR at 10db 0.8 0.6 0.4 0.2 0 Mag -0.2 -0.4 -0.6 -0.8 -1 -1.2 0 0.5 1 1.5 2 Samples Paul Coffey Project Presentation 2.5 4 x 10 Testing A blind test was carried out involving ten student in the Electronics Department Consisted of number of different samples at different levels of SNR Results from this test showed at low levels of SNR the two filters are very close Then at higher levels of SNR the Wiener Filter is preferred Paul Coffey Project Presentation Conclusion Speech Enhancement techniques are a very important part of Communication Systems The two techniques, Spectral Subtraction and the Wiener Filter were implemented From the testing the Wiener Filter was the preferred choice over the Spectral Subtraction Paul Coffey Project Presentation Questions? Paul Coffey Project Presentation