def process(wav_filename, expected_peak_freq):
sample_rate, data_unnormalized = scipy.io.wavfile.read(wav_filename)
data = (numpy.array(data_unnormalized, dtype=numpy.float64)
/ float(numpy.iinfo(data_unnormalized.dtype).max))
# expermental: remove some data
#data = data[0.2*len(data):0.5*len(data)]
if SHORT:
data = data[:SHORT]
freqs, fft, phase, fft_length = get_freqs_fft(data, sample_rate)
fft_db = dsp.amplitude2db(fft)
if NORMALIZE_DB:
fft_db -= max(fft_db)
filename = "%s-freqs.txt" % (os.path.splitext(wav_filename)[0] )
numpy.savetxt( filename, numpy.vstack( (freqs, fft_db)).transpose() )
low_freq = expected_peak_freq - FREQ_WIDTH
high_freq = expected_peak_freq + FREQ_WIDTH
low_bin = hertz2bin(low_freq, sample_rate, fft_length)
high_bin = hertz2bin(high_freq, sample_rate, fft_length)
freqs = freqs [low_bin:high_bin]
fft_db = fft_db[low_bin:high_bin]
phase = phase [low_bin:high_bin]
return freqs, fft_db, phase
def process(filename):
sample_rate, data_unnormalized = scipy.io.wavfile.read(filename)
data = numpy.array(data_unnormalized, dtype=numpy.float64)
data = data_unnormalized / float(numpy.iinfo(data_unnormalized.dtype).max)
#data = scipy.signal.decimate(data, 2)
#sample_rate /= 2.0
cutoff = 20.0 / (sample_rate/2.0)
pylab.plot(data)
b, a = scipy.signal.butter(4, cutoff, btype='highpass')
data = scipy.signal.lfilter(b, a, data)
#pylab.plot(data)
#pylab.ylim([-1,1])
#pylab.show()
#data = numpy.array(data, dtype=numpy.int32)
data_abs = abs(data)
ind = data_abs.argmax()
xs = numpy.arange(0, len(data))
orientation = float(data[ind])
for f in xrange(ind, ind-1000, -1):
# is a zero crossing
#print orientation, data[f]
if orientation * data[f] <= 0:
zero = f
break
#print ind, zero
save = data [zero : zero+4096]
xs_save = xs[zero : zero+4096]
#print filename, ind, zero
#pylab.plot(xs[:ind], data[:ind])
#pylab.plot(xs_save[:500], save[:500])
#pylab.show()
out_filename = os.path.join(split_dir,
os.path.basename(filename))
saveints = numpy.array( save * float((2**16)-1), dtype=numpy.int16)
scipy.io.wavfile.write(out_filename, sample_rate, saveints)
#fft = scipy.fftpack.fft(save / float(numpy.iinfo(save.dtype).max))
fft = scipy.fftpack.fft(save)
fftabs = abs(fft[:len(fft)/2]) / len(save)
freqs = numpy.array([
float(i) * sample_rate / len(fft)
for i in xrange(len(fftabs))
])
fftdb = dsp.amplitude2db(fftabs)
data = numpy.vstack( (freqs, fftdb) ).transpose()
numpy.savetxt(out_filename[:-4] + ".freqs.txt", data)
data = numpy.vstack( (
numpy.arange(0, len(save))/44100.0, save * float(2**31-1) ) ).transpose()
numpy.savetxt(out_filename[:-4] + ".time.txt", data)
def fit_lorentzian(freqs, fft_db, plot=False, add_plot=False):
w = freqs
y = dsp.db2amplitude(fft_db)
#y = fft_db
#w = numpy.array( [
# bin2hertz(bin_low+i) for i, a in enumerate(peak_area)
# ] )
#y = numpy.array(peak_area)
#noise_floor = scipy.stats.gmean(y)
noise_floor = min(y)
mag_est = max(y) - noise_floor
initial_guess = numpy.array([1.0, w.mean(), mag_est,
noise_floor
])
#print "initial:", initial_guess
p, pcov = scipy.optimize.curve_fit(
lorentzian,
w, y,
initial_guess,
)
#print "fit:", p
predicted = lorentzian(w, *p)
#if plot:
# #pylab.plot(w, dsp.amplitude2db(y), label="FFT")
# #pylab.plot(w, dsp.amplitude2db(predicted),
# # label="Lorentzian fit")
# pylab.plot(w, y, label="FFT")
# pylab.plot(w, predicted,
# label="Lorentzian fit")
# pylab.show()
#
ss_err = ((y - predicted)**2).sum()
ss_tot = ((y-y.mean())**2).sum()
rsquared = 1.-(ss_err/ss_tot)
#try:
# variance = pcov[0][0]
#except:
# print "bail from variance"
# return None, 0, 0, 0
variance = 0
freq = p[1]
mag = p[2]
decay = 2*numpy.pi / abs(p[0])
#print "Fit first:\t%.1f\t%.2g\t%.3g" % (
print "%.2f\t%.3g\t%.4g" % (
freq, mag, decay)
#if rsquared < LONG_LORENTZIAN_RSQUARED_MIN:
# print "bail from rsquared", rsquared
# return freq, 0, 0, 0
if plot:
logplot = True
pylab.figure()
if logplot:
pylab.plot(w, dsp.amplitude2db(y), label="FFT")
pylab.plot(w, dsp.amplitude2db(predicted),
label="Lorentzian fit")
#pylab.semilogy(w, resi, label="residuals")
else:
pylab.plot(w, y, label="FFT")
pylab.plot(w, predicted, label="Lorentzian fit")
#pylab.plot(w, resi, label="residuals")
#pylab.xlabel("frequency relative to w_n")
pylab.xlabel("frequency")
pylab.ylabel("power")
pylab.legend()
pylab.show()
if add_plot:
low = min(w_eval)
high = max(w_eval)
pylab.axvspan(low, high, color='c', alpha=0.3,
label="stiff")
pylab.plot(w_eval, numpy.log(predicted),
#pylab.plot(w_eval, (predicted),
label="Lorentzian fit",
color="red")
return freq, decay, rsquared, variance
def pluck_force(violin, st, force, finger, plot=False, write=False):
#violin = artifastring_instrument.ArtifastringInstrument(inst, instnum)
#wavfile = monowav.MonoWav("artifastring-test.wav")
violin.reset()
violin.finger(st, finger)
violin.pluck(st, 0.2, force)
def hop():
buf = numpy.empty(HOPSIZE, dtype=numpy.int16)
forces = numpy.empty(FORCE_SIZE, dtype=numpy.int16)
violin.wait_samples_forces_python(buf, forces)
string_array = numpy.zeros(4*HOPSIZE, dtype=numpy.float32)
string_array_size = violin.get_string_buffer(st, string_array)
string_array = string_array[:string_array_size]
#pylab.plot(string_array)
#pylab.show()
buf_rms = numpy.sqrt(numpy.mean(numpy.array(buf,
numpy.float64)**2))
sa_rms = numpy.sqrt(numpy.mean(numpy.array(string_array,
dtype=numpy.float64)**2))
buf_ss = numpy.sum(numpy.array(buf,
numpy.float64)**2)
sa_ss = numpy.sum(numpy.array(string_array,
dtype=numpy.float64)**2)
return buf_rms, sa_rms, buf_ss, sa_ss
dh = float(HOPSIZE) / artifastring_instrument.ARTIFASTRING_INSTRUMENT_SAMPLE_RATE
BUFS = 1000
dhs = numpy.arange(0, BUFS) * dh
buf = numpy.zeros(BUFS)
sa = numpy.zeros(BUFS)
buf_sss = numpy.zeros(BUFS)
sa_sss = numpy.zeros(BUFS)
for i in range(BUFS):
buf_this, string_array_this, buf_ss, sa_ss = hop()
buf[i] = buf_this
sa[i] = string_array_this
buf_sss[i] = buf_ss
sa_sss[i] = sa_ss
buf_db = dsp.amplitude2db(buf)
sa_db = dsp.amplitude2db(sa)
cutoff_hop = 0
for i in range(BUFS):
if buf_db[i] < CUTOFF_DB_FINAL_AUDIO:
cutoff_hop = i
break
print "cutoff time:", cutoff_hop*dh
cutoff_internal_audio = sa_sss[cutoff_hop]
cutoff_internal_audio_db = sa_db[cutoff_hop]
#print "Cutoff internal audio:", cutoff_internal_audio
if write:
numpy.savetxt("instrument-%i-%.3f-db.txt" % (
st,finger),
numpy.vstack( (
dhs, buf_db
)).transpose())
numpy.savetxt("string-%i-%.3f-db.txt" % (
st,finger),
numpy.vstack( (
dhs, sa_db
)).transpose())
numpy.savetxt("cutoff-%i-%.3f.txt" % (
st,finger),
numpy.array([
[0,
cutoff_internal_audio_db],
[BUFS*dh,
cutoff_internal_audio_db]
])
)
if plot:
pylab.subplot(211)
pylab.title("Final audio")
pylab.plot(dhs, buf_db)
pylab.axhline(CUTOFF_DB_FINAL_AUDIO)
pylab.subplot(212)
pylab.title("String audio")
pylab.plot(dhs, sa_db, '.-')
pylab.axhline(cutoff_internal_audio_db)
pylab.show()
return cutoff_internal_audio
def fit_area(freqs, fft_db, plot=False, add_plot=False):
w = freqs
y = dsp.db2amplitude(fft_db)
#y = fft_db
#w = numpy.array( [
# bin2hertz(bin_low+i) for i, a in enumerate(peak_area)
# ] )
#y = numpy.array(peak_area)
global noise_floor
#noise_floor = scipy.stats.gmean(y)
noise_floor = min(y)
mag_est = max(y) - noise_floor
initial_guess = numpy.array([1.0, w.mean(), mag_est
])
print "initial:", initial_guess
p, pcov = scipy.optimize.curve_fit(
lorentzian,
w, y,
initial_guess,
)
print "fit:", p
freq = p[1]
mag = p[2]
decay = 2*numpy.pi / abs(p[0])
print "Fit only:\t%.1f\t%.2g\t%.3g" % (
freq, mag, decay)
predicted = lorentzian(w, *p)
rem = y - predicted
#if False:
if True:
pylab.plot(w, dsp.amplitude2db(y), label="FFT")
pylab.plot(w, dsp.amplitude2db(predicted),
label="Lorentzian fit")
pylab.plot(w, dsp.amplitude2db(rem),
label="residual")
#pylab.plot(w, y, label="FFT")
#pylab.plot(w, predicted,
# label="Lorentzian fit")
pylab.show()
#double_guess = numpy.array([ p[0], p[1], p[2], p[0], p[1],
# p[2]/100.0])
double_guess = numpy.array([ p[0], p[1], p[2],
1.2, 96.3, 200.0])
#print "double guess:"
#print double_guess
p, pcov = scipy.optimize.curve_fit(
lorentzian_two,
w, y,
double_guess
)
print "fit two:"
print p
predicted = lorentzian_two(w, *p)
#logplot = False
#pylab.figure()
#if logplot:
# pylab.semilogy(w, y, '.', label="FFT")
# pylab.semilogy(w_eval, predicted, label="Lorentzian fit")
# pylab.show()
#else:
# pylab.plot(w, y, '.', label="FFT")
# pylab.plot(w_eval, predicted, label="Lorentzian fit")
# pylab.show()
ss_err = ((y - predicted)**2).sum()
ss_tot = ((y-y.mean())**2).sum()
rsquared = 1.-(ss_err/ss_tot)
#try:
# variance = pcov[0][0]
#except:
# print "bail from variance"
# return None, 0, 0, 0
variance = 0
freq = p[1]
mag = p[2]
decay = 2*numpy.pi / abs(p[0])
print "Fit first:\t%.1f\t%.2g\t%.3g" % (
freq, mag, decay)
print "Fit second :\t%.1f\t%.2g\t%.3g" % (
p[4], p[5], 2*numpy.pi / abs(p[3]))
#if rsquared < LONG_LORENTZIAN_RSQUARED_MIN:
# print "bail from rsquared", rsquared
# return freq, 0, 0, 0
rem = y - predicted
if plot:
logplot = True
pylab.figure()
if logplot:
pylab.plot(w, dsp.amplitude2db(y), label="FFT")
pylab.plot(w, dsp.amplitude2db(predicted),
label="Lorentzian fit")
pylab.plot(w, dsp.amplitude2db(rem),
label="residual")
#pylab.semilogy(w, resi, label="residuals")
else:
pylab.plot(w, y, label="FFT")
pylab.plot(w, predicted, label="Lorentzian fit")
#pylab.plot(w, resi, label="residuals")
#pylab.xlabel("frequency relative to w_n")
pylab.xlabel("frequency")
pylab.ylabel("power")
pylab.legend()
if add_plot:
low = min(w_eval)
high = max(w_eval)
pylab.axvspan(low, high, color='c', alpha=0.3,
label="stiff")
pylab.plot(w_eval, numpy.log(predicted),
#pylab.plot(w_eval, (predicted),
label="Lorentzian fit",
color="red")
return freq, decay, rsquared, variance
