03-03-2011, 09:50 AM
PRESENTED BY:
K.AKILA
N.DEEPTHA
D.NARMADA
final ppt.ppt (Size: 1.18 MB / Downloads: 71)
NOISE REDUCTION IN SPEECH
• Noise is an unwanted disturbance superimposed on a
useful signal, which tends to obscure its information
content.
• The background noise is the most common factor
degrading the quality and intelligibility of speech .
• The noise reduction technique intends to lower the
noise level without affecting the quality of speech signal.
OBJECTIVE
• To achieve intelligibility, naturalness and overall
perceptual quality of the speech signal.
• To improve the signal strength.
NOISE DATA
Factory Noise
HF Channel Noise
Babble Noise
White Noise
Pink Noise
Car Interior Noise
Cockpit Noise
N.S - Noisy Speech
A.C - Approximation Coefficients
D.C - Detail Coefficients
ICA - Independent Component Analysis
ZCR - Zero Crossing Rate
N.R.S - Noise Reduced Speech
STANDARD DATABASE
SPEECH SIGNAL :
TEXT : SHE HAD YOUR DARK SUIT IN GREASY WASH
WATER ALL YEAR
SAMPLE LENGTH : 63488 samples
SAMPLING FREQUENCY :16000Hz
VARIANCE :E(X2)-(E(X))2
Variance of the SPEECH SIGNAL(in dB) : -74.5232
ORIGINAL SPEECH SIGNAL
FREQUENCIES IN THE SPEECH SIGNAL
FREQUENCIES IN THE NOISE
NOISY SPEECH SIGNAL
FREQUENCIES IN THE NOISY SPEECH
VARIANCE :E(X2)-(E(X))2
Variance of the SPEECH SIGNAL(in dB) : -74.5232
Variance of the NOISE(in dB) : -106.1482
Variance of the NOISY SPEECH(in dB) : -74.1012
FRAMING
NON-OVERLAPPING FRAMES OF 16msec DURATION.
NUMBER OF FRAMES :248
WAVELET DECOMPOSITION
WAVELET USED - DAUBECHIES WAVELET db1
DECOMPOSITION LEVEL - LEVEL 2
APPROXIMATION COEFFICIENTS :
Correspond to speech combined with low frequency noise in the frequency band (0-2000Hz).
DETAIL COEFFICIENTS :
Correspond to the contribution due to speech in the frequency band (2000-4000Hz).
WAVELET RECONSTRUCTION
APPROXIMATION COEFFICIENTS AND DETAIL
COEFFICIENTS are UPSAMPLED.
RECONSTRUCTED SIGNAL
Variance of the RECONSTRUCTED
SIGNAL in(0-2000 Hz) (in dB) : -76.4146
EXTRACTION OF NOISE FROM NOISY SPEECH
To analyze the noise frequencies, we need to extract the noise from noisy speech.
Wavelet transform of LEVEL 2 is used to extract the speech in low frequency band (0-2000 Hz) and to find the starting frame.
A Modified probability density function is calculated.
A negative wavelet entropy computed using modified probability density function is used for determining the starting frame.
MODIFIED PROBABILITY DENSITY FUNCTION
i-frame number
j-sample in a frame
Wi(j) –wavelet energy of the jth sample in ith frame
K-positive constant=1.8
NEGATIVE ENTROPY CALCULATION
The negative wavelet entropy is given as
E(i) = Pi(j)*log(Pi(j))
Pi(j) -Modified probability function
The negative wavelet entropy is more negative for noise and silence frames and less negative for speech frames.
Starting frame is determined from the entropy comparison
ENTROPY PLOT
DETERMINING THE STARTING FRAME
A THRESHOLD is set
THRESHOLD=(MEAN(entropy)+MEDIAN(entropy))/2
THRESHOLD= -74.32
STARTING FRAME :46
The ending frame is also determined.
END FRAME : 234
BOUNDARY PLOT
CONCLUSION
Variance of the NOISY SPEECH(in dB)= -74.1012
Variance of the RECONSTRUCTED = -76.4146
SIGNAL from wavelet decomposition (in dB)
There is an improvement in the intelligibility of the noisy speech which is proved by hearing the speech at each stage.