You are on page 1of 16

Signal processing

Signals, Noise and Information


• A signal is the variation of a quantity [Vaseghi-2008] that conveys
information regarding one or more attributes of the source
• A signal is a means of conveying information regarding the past,
the current or the future states of a variable
Communication and signal processing system

system T[·] for transformation of the information into variation of a signal x(t)

communication channel h[·] for modelling the propagation of the signal


from the transmitter to the receiver

additive channel and background noise n(t)


Communication and signal processing system

x(t) = T[I(t)]

y(t) = h[x(t)] + n(t)


Noise reduction
• Amrulloh et al. – 2015: Cough
• HPF: High pass filter
• Cut off frequency: 10Hz
• PSS: Power spectral subtraction
• Reduce the Gaussian noisse
• DeepCough [Amoh - 2015]
• Frame admission control [Lu-2015]: For each 16-frame (64 ms) window,
the RMS energy is calculated and compared with a predetermined threshold.
Windows with low energy are assumed to be silence or ambient sound, and
they are discarded. High energy windows are "admitted" and undergo further
processing. For
Frame admission control
• Frame may contain audio content that is not interesting (e.g., white noise) or is
not able to be classified (e.g., silence or insufficient amount of the signal is
captured)
• Based on energy level and spectral entropy
• Low energy  silences or undersirable context  rmsthreshold
• Computed by RMS (Root mean square) of the amplitude of the audio content it contains.
• Flat spectrum (silence or white noise)  high entropy  entropythreshold
• apply a hanning window to the frame, which suppresses the frame boundaries and thus
reduces the known effect of Fast Fourier Transform (FFT) spectral leakage
• calculate the FFT spectrum of the frame
• normalize the spectrum, treat it as a probability density function
• Admitted frames: RMS > rmsthreshold or entropy < entropythreshold
• Eliminate conversations and music
• Frame with short period of time (five seconds)
Frame admission control
• Eliminate conversations and music
• Frame with short period of time (five seconds)
RNNoise – Recurrent Neural Networks Noise
• Recurrent neural networks (RNN) are very important here because
they make it possible to model time sequences instead of just
considering input and output frames independently. This is especially
important for noise suppression because we need time to get a good
estimate of the noise.
Demo session
High/Low pass filter [Amrulloh, et al. - ]
• Sound s with noise
• Convert the recordings sound to discrete signal data
• Create hpf as HPF fourth order Butterworth in Matlab, F_cut_off Hz
• Filter the signal by the created HPF, get b
[z,p,k] = butter(4,f_cut_off/(fs/2),str)
[sos, g] = zp2sos(z, p, k);
freqz(sos, 2^16);
b= filtfilt(sos, g, data);
audiowrite('b.wav',b,fs);
str = high or low
Low Pass Filter []

• Fs = xxx; % Sampling Frequency (Hz)


• Fn = Fs/2; % Nyquist Frequency
• Fco = xxx; % Passband (Cutoff) Frequency
• Fsb = 30; % Stopband Frequency
• Rp = 1; % Passband Ripple (dB)
• Rs = 10; % Stopband Ripple (dB)
• [n,Wn] = buttord(Fco/Fn, Fsb/Fn, Rp, Rs); % Filter Order & Wco
• [b,a] = butter(n,Wn); % Lowpass Is Default Design
• [sos,g] = tf2sos(b,a); % Second-Order-Section For Stability
• Data_LPF = filtfilt(sos, g, data);
Cough: HPF
Original cough sound

After HPF ; Cut-off fre: 1000Hz


Crackle: HPF Original crackle sound

After HPF; Cut-off fre: 1500Hz


Breath

HPF

LPF
Breath

HPF

LPF
Amrulloh, et al.
• Select cough s with noise
• Estimate noise power spectral density
• the minimum power level in any given frequency bin when observed over a sufficiently
long period represents the power in the noise signal [Martin-2001].
• Create pss as PSS filter in Matlab
• Filter b by pss to get c
• Listen and compare s and c

You might also like