peaker = loudia.PeakDetectionComplex(
maxPeakCount, loudia.PeakDetectionComplex.BYMAGNITUDE, minPeakWidth)
peakInterp = loudia.PeakInterpolationComplex()
tracker = loudia.PeakTracking(maxTrajCount, maxFreqBinChange, silentFrames)
filt = loudia.BandFilter()
trajsLocs = []
trajsMags = []
specs = []
filtereds = []
for frame in stream:
samples = frame
fft = ffter.process(windower.process(frame))[0, :]
spec = loudia.magToDb(abs(fft))[0, :plotSize]
if set(['phase', 'peak_phases']) | all_processes:
phase = scipy.angle(fft)
if set(['peak_mags', 'peak_phases']) | all_processes:
peakLocs, peakMags, peakPhases = peaker.process(fft)
peakiLocs, peakiMags, peakiPhases = peakInterp.process(
fft, peakLocs, peakMags, peakPhases)
trajLocs, trajMags = tracker.process(fft, peakiLocs, peakiMags)
filtered = scipy.reshape(samples, (samples.shape[0], 1))
d1 = loudia.freqz(coeffsB1, coeffsA, w)
d2 = loudia.freqz(coeffsB2, coeffsA, w)
d3 = loudia.freqz(coeffsB3, coeffsA, w)
d4 = loudia.freqz(coeffsB4, coeffsA, w)
d = d1 * d2 * d3 * d4
# Plot the frequency response
import pylab
subplots = 2 if plotAngle else 1;
pylab.subplot(subplots,1,1)
if plotColor:
pylab.plot(w[npoints/2:], loudia.magToDb(abs(d[npoints/2:,:])))
else:
pylab.plot(w[npoints/2:], loudia.magToDb(abs(d[npoints/2:,:])), c = 'black')
pylab.hold(True)
suma = abs(d[npoints/2:,:]).sum(axis=1)
suma.resize((suma.shape[0], 1))
pylab.plot(w[npoints/2:], loudia.magToDb(suma), c = 'red')
ax = pylab.gca()
# Show half of the spectrum
ax.set_xlim([0, scipy.pi])
# Set the ticks units to radians per second
#!/usr/bin/env python import loudia import scipy import scipy.signal # Test the magToDb <-> dbToMag i = scipy.arange(1, 7, dtype = 'f4') a = loudia.magToDb(i) o = loudia.dbToMag(a) print 'magToDb/dbToMag:', scipy.allclose(i, o) # Test the transform of windows transf = loudia.hammingTransform(24, 10, 1024, 4096) # Test the poly function a = scipy.array([1, 2, 3, 4, 5]) rr = loudia.poly( a ) rs = scipy.poly( a ) print 'poly:', scipy.allclose(rr, rs) a = scipy.array([2, 0]) rr = loudia.poly( a ) rs = scipy.poly( a[:-1] ) print 'poly, plus zero:', scipy.allclose(rr[0,:-1], rs)
import scipy
from common import *
import pylab
import loudia
import sys
filename = sys.argv[1]
frameSize = 256
frameStep = 256
fftSize = frameSize * (2**3)
analysisLimit = 1000.0
plotSize = fftSize / 4
stream, sampleRate, nframes, nchannels, loader = get_framer_audio(filename, frameSize, frameStep)
windower = loudia.Window( frameSize, loudia.Window.HAMMING )
ffter = loudia.FFT( fftSize )
pylab.figure()
for ind, frame in enumerate(stream):
if ind == 34:
fft = ffter.process( windower.process( frame ) )
spec = loudia.magToDb( abs( fft ) )
pylab.plot( spec[0, :plotSize] )
pylab.show()
peaker = loudia.PeakDetectionComplex( maxPeakCount, loudia.PeakDetectionComplex.BYMAGNITUDE, minPeakWidth )
peakInterp = loudia.PeakInterpolationComplex( )
tracker = loudia.PeakTracking( maxTrajCount, maxFreqBinChange, silentFrames )
peakSynth = loudia.PeakSynthesize( frameSize/6, fftSize, windowType )
trajsLocs = []
trajsMags = []
specs = []
specsSynth = []
specsResid = []
specsMagsResid = []
for frame in stream:
samples = frame
fft = ffter.process( windower.process( frame ) )[0, :plotSize]
spec = loudia.magToDb( abs( fft ) )[0, :plotSize]
if set(['phase', 'peak_phases']) | all_processes:
phase = scipy.angle( fft )
if set(['peak_mags', 'peak_phases']) | all_processes:
fft = scipy.reshape(fft, (1, plotSize))
peakLocs, peakMags, peakPhases = peaker.process( fft )
peakiLocs, peakiMags, peakiPhases = peakInterp.process( fft,
peakLocs,
peakMags,
peakPhases )
trajLocs, trajMags = tracker.process( fft,
errors = []
npoints = 1024
w = scipy.arange(0, scipy.pi, scipy.pi/(fftSize/2. + 1))
b = scipy.array([[1]])
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
fft = ffter.process( windower.process( frame ) )[0, :]
spec = loudia.magToDb( abs( fft ) )
lpcCoeffs, reflection, error = lpc.process( frame )
lpcResidual = lpcr.process( frame, lpcCoeffs )
freqResp = loudia.magToDb( abs( loudia.freqz( b * scipy.sqrt( abs( error[0] ) ), lpcCoeffs.T, w ) ).T )
if interactivePlot:
pylab.subplot( 211 )
pylab.hold( False )
pylab.plot( frame )
#pylab.hold( True )
#pylab.plot( lpcResidual[0,:] )
pylab.subplot( 212 )
specs = []
wspecs = []
pitches = []
saliencies = []
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
fft = ffter.process(windower.process(frame))
spec = loudia.magToDb(abs(fft))
wspec = whitening.process(spec)
pitch, saliency = pitchACF.process(wspec)
if interactivePlot:
pylab.subplot(211)
pylab.hold(False)
pylab.plot(wspec[0, :plotSize])
acorred = acorr.process(wspec)
pylab.subplot(212)
pylab.hold(False)
pylab.plot(acorred[0, :plotSize], label='Noise Suppressed Spectrum')
pylab.hold(True)
saliencies = []
freqss = []
frameNumber = 200
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the frequency likelihood')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
samples = frame
fft = ffter.process( windower.process( frame ) )
spec = loudia.magToDb( abs( fft ) )
wspec = whitening.process( spec )
#wspec = spec
pitch, saliency, freqs = pitchInverseProblem.process( spec )
frameNumber -= 1
if interactivePlot:
if frameNumber == 0:
maxFreq = freqs[0,:].argmax()
pylab.subplot(211)
pylab.hold(True)
pylab.plot( spec[0, :plotSize] )
pylab.plot(loudia.magToDb(a)[:, maxFreq] + 40)
wspecs = []
pitches = []
saliencies = []
freqss = []
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
samples = frame
fft = ffter.process( windower.process( frame ) )[0, :plotSize]
spec = loudia.magToDb( abs( fft ) )[0, :plotSize]
wspec = whitening.process( spec )
pitch, saliency, freqs = pitchInverseProblem.process( wspec )
if interactivePlot:
pylab.subplot(211)
pylab.hold(False)
pylab.plot( wspec[0, :plotSize] )
pylab.subplot(212)
pylab.hold(False)
pylab.plot(freqs[0,:plotSize], label = 'Noise Suppressed Spectrum')
#pylab.hold(True)
#pylab.stem( pitch/sampleRate*fftSize, saliency )
pylab.figure()
for band in range(m.bands()):
pylab.hold(True)
weight = m.bandWeights( band ).T
start = starts[band]
suma[start:start+weight.shape[1]] = suma[start:start+weight.shape[1]] + weight.T
x = scipy.arange(start-1, start + weight.shape[1]+1)
y = [0] + list(weight[0,:]) + [0]
y = scipy.array(y, dtype = 'f4')
y.resize((y.shape[0], 1))
pylab.plot(x, loudia.magToDb(y), color = 'black')
ax = pylab.gca()
# Show half of the spectrum
ax.set_xlim([0, spectrumSize / 2])
ax.set_ylim([0.0, 1.1])
# Set the ticks units to radians per second
ticks = ax.get_xticks()
ax.set_xticklabels(['%.2f' % (float(tick) / spectrumSize) for tick in ticks])
# Set the title and labels
pylab.title('Magnitude of the Frequency Response of a \n Mel Bands implementation')
pylab.xlabel('Normalized Frequency')
pylab.ylabel('|H(w)| (no unit)')
#!/usr/bin/env python import loudia import scipy import scipy.signal # Test the magToDb <-> dbToMag i = scipy.arange(1, 7, dtype='f4') a = loudia.magToDb(i) o = loudia.dbToMag(a) print 'magToDb/dbToMag:', scipy.allclose(i, o) # Test the transform of windows transf = loudia.hammingTransform(24, 10, 1024, 4096) # Test the poly function a = scipy.array([1, 2, 3, 4, 5]) rr = loudia.poly(a) rs = scipy.poly(a) print 'poly:', scipy.allclose(rr, rs) a = scipy.array([2, 0]) rr = loudia.poly(a) rs = scipy.poly(a[:-1]) print 'poly, plus zero:', scipy.allclose(rr[0, :-1], rs)
windower = loudia.Window(frameSize, loudia.Window.HAMMING)
ffter = loudia.FFT(fftSize)
supprnoise = loudia.SpectralNoiseSuppression(fftSize, 50.0, 6000.0, sampleRate)
specs = []
results = []
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
fft = ffter.process(windower.process(frame))
spec = loudia.magToDb(abs(fft))
noise, result = supprnoise.process(abs(fft))
if interactivePlot:
pylab.subplot(211)
pylab.hold(False)
specdb = loudia.magToDb(spec)
noisedb = loudia.magToDb(noise)
pylab.plot(specdb[0, :], label='Spectrum')
pylab.hold(True)
pylab.plot(noisedb[0, :], label='Noise')
resultdb = loudia.magToDb(result)
d1 = loudia.freqz(coeffsB1, coeffsA, w)
d2 = loudia.freqz(coeffsB2, coeffsA, w)
d3 = loudia.freqz(coeffsB3, coeffsA, w)
d4 = loudia.freqz(coeffsB4, coeffsA, w)
d = d1 * d2 * d3 * d4
# Plot the frequency response
import pylab
subplots = 2 if plotAngle else 1
pylab.subplot(subplots, 1, 1)
if plotColor:
pylab.plot(w[npoints / 2:], loudia.magToDb(abs(d[npoints / 2:, :])))
else:
pylab.plot(w[npoints / 2:],
loudia.magToDb(abs(d[npoints / 2:, :])),
c='black')
pylab.hold(True)
suma = abs(d[npoints / 2:, :]).sum(axis=1)
suma.resize((suma.shape[0], 1))
pylab.plot(w[npoints / 2:], loudia.magToDb(suma), c='red')
ax = pylab.gca()
# Show half of the spectrum
ax.set_xlim([0, scipy.pi])
windower = loudia.Window( frameSize, loudia.Window.HAMMING )
ffter = loudia.FFT( fftSize )
supprnoise = loudia.SpectralNoiseSuppression(fftSize, 50.0, 6000.0, sampleRate)
specs = []
results = []
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
fft = ffter.process( windower.process( frame ) )
spec = loudia.magToDb( abs( fft ) )
noise, result = supprnoise.process( abs( fft ) )
if interactivePlot:
pylab.subplot( 211 )
pylab.hold( False )
specdb = loudia.magToDb( spec )
noisedb = loudia.magToDb( noise )
pylab.plot( specdb[0,:], label = 'Spectrum' )
pylab.hold( True )
pylab.plot( noisedb[0,:], label = 'Noise' )
resultdb = loudia.magToDb( result )
errors = []
npoints = 1024
w = scipy.arange(0, scipy.pi, scipy.pi / (fftSize / 2. + 1))
b = scipy.array([[1]])
if interactivePlot:
pylab.ion()
pylab.figure()
pylab.title('Interactive plot of the FFT vs LPC frequency response')
pylab.gca().set_ylim([-100, 40])
pylab.gca().set_autoscale_on(False)
for frame in stream:
fft = ffter.process(windower.process(frame))[0, :]
spec = loudia.magToDb(abs(fft))
lpcCoeffs, reflection, error = lpc.process(frame)
lpcResidual = lpcr.process(frame, lpcCoeffs)
freqResp = loudia.magToDb(
abs(loudia.freqz(b * scipy.sqrt(abs(error[0])), lpcCoeffs.T, w)).T)
if interactivePlot:
pylab.subplot(211)
pylab.hold(False)
pylab.plot(frame)
#pylab.hold( True )
#pylab.plot( lpcResidual[0,:] )
predicted = 2*angsUnwrapped[1,:] - angsUnwrapped[0,:]
error_linearity = abs(predicted - angsUnwrapped[2,:])
error_slope = abs((angsUnwrapped[1,:] + 2 * scipy.pi * scipy.arange(angsUnwrapped.shape[1])/(angsUnwrapped.shape[1]-1) * miniHopSize/2.0) - angsUnwrapped[2,:]) % (2*scipy.pi)
error_linearity = mapError(error_linearity)
error_slope = mapError(error_slope)
errors.append(error_linearity + error_slope)
# The prediction error will be the inverse of the sinusoidity
if frameCount == 40:
pylab.figure()
pylab.subplot(311)
pylab.plot(loudia.magToDb(cAbs)[0, :plotSize])
pylab.subplot(312)
pylab.plot(angsUnwrapped[-1, :plotSize])
pylab.subplot(313)
pylab.plot(error_linearity[:plotSize], label='linearity')
pylab.hold(True)
pylab.plot(error_slope[:plotSize], label='slope', c='g')
frameCount += 1
cursor += hopSize
errors = scipy.array( errors )
import numpy
def smooth(x,window_len=11,window='hanning'):
