def signal_energy(signal, fs, fl=0.05, fh=0.02):
X = stft(signal, fs, fl, fh)
#pylab.figure();
#pylab.imshow(scipy.absolute(X.T), origin='lower', aspect='auto',
# interpolation='nearest')
#pylab.specgram(signal, NFFT=44100*fs, noverlap=44100*fh);
#pylab.show();
mag = scipy.absolute(X)
energy = scipy.sum(mag, 1)
return energy
def signal_energy(signal, fs, fl = 0.05, fh=0.02):
X = stft(signal, fs, fl,fh)
#pylab.figure();
#pylab.imshow(scipy.absolute(X.T), origin='lower', aspect='auto',
# interpolation='nearest')
#pylab.specgram(signal, NFFT=44100*fs, noverlap=44100*fh);
#pylab.show();
mag = scipy.absolute(X);
energy = scipy.sum(mag, 1)
return energy
def computeSpectrogram(self, timeseries, nFFT, nHop, dt, windowFn, fRange):
'''
Compute STFT spectrogram of data in timeseries. Expected numpy.array shape
(nSamples,nSignals)
Returns: spectra, (nTimes,nFreq,nSignals)
'''
data = timeseries.getData()
fVec = self.getFreqVec()
if fRange is not None:
mask = np.logical_and(fVec >= fRange[0], fVec <= fRange[1])
else:
# Create an all true array
mask = np.ones((len(fVec), 1)) == 1
mask = np.nonzero(mask)
mask = mask[0]
# Loop over signals (2nd dimension of )
Nsignals = timeseries.numSignals()
for i in range(Nsignals):
x = data[:, i, np.newaxis]
X = spectrogram.stft(x,
nFFT,
nHop,
transform=np.fft.fft,
win=windowFn,
zp_back=0,
zp_front=0)
if i == 0:
spectra = np.zeros((X.shape[0], len(mask), Nsignals),
dtype=complex)
spectra[:, :, i] = X[:, mask]
return spectra
def transform(audio_data, save_image_path, nFFT=256, overlap=0.75):
'''audio_data: signals to convert
save_image_path: path to store the image file'''
# spectrogram
freq_data = stft(audio_data, nFFT, overlap)
freq_data = np.maximum(np.abs(freq_data),
np.max(np.abs(freq_data)) / 10000)
log_freq_data = 20. * np.log10(freq_data / 1e-4)
N_samples = log_freq_data.shape[0]
# log_freq_data = np.maximum(log_freq_data, max_m - 70)
# print(np.max(np.max(log_freq_data)))
# print(np.min(np.min(log_freq_data)))
log_freq_data = np.round(log_freq_data)
log_freq_data = np.transpose(log_freq_data)
# ipdb.set_trace()
assert np.max(np.max(log_freq_data)) < 256, 'spectrogram value too large'
# save the image
spec_imag = Image.fromarray(log_freq_data)
spec_imag = spec_imag.convert('RGB')
spec_imag.save(save_image_path)
return N_samples
def find_words(data, fs, num, **args):
min_silence_length = args["min_silence_length"] if "min_silence_length" in args else 5
_plot = args["plot"] if "plot" in args else False
_debug = False
frml = 0.03
frmh = 0.02
X = stft(data, fs, frml, frmh)
magnitude = np.absolute(X)
signal_energy = np.sum(magnitude, 1)
if "thresh" in args.keys():
thresh = args["thresh"]
else:
thresh = np.mean(signal_energy) - np.min(signal_energy)
en_th = (signal_energy > thresh).astype(np.int8)
rising_edge = np.diff( np.hstack(([0], en_th)) )
falling_edge = np.diff( np.hstack((en_th, [0])) )
(stp,) = np.nonzero(rising_edge > 0)
(edp,) = np.nonzero(falling_edge < 0)
if len(stp) != len(edp):
raise ValueError("Wrong threshold value")
"""Remove too short silence periods"""
silence = stp[1:] - edp[:-1]
(too_short,) = np.nonzero(silence <= min_silence_length)
stp = np.delete(stp, too_short + 1)
edp = np.delete(edp, too_short)
nonsilence = edp - stp
shortest = nonsilence.argsort()
if len(nonsilence) < num:
raise ValueError("You say less words than you ask me to find or you speak too fast")
""" Remove too short nonsilence """
stp = np.delete(stp, shortest[:-num])
edp = np.delete(edp, shortest[:-num])
pos = lambda a: fs*frmh*a
if _plot:
pylab.subplot(2,1,1)
etime = np.array(range(len(signal_energy)), dtype=float)*frmh
pylab.plot(etime,signal_energy/1e6)
pylab.axhline(thresh/1e6)
[ pylab.axvline(x*frmh, color="red") for x in stp ]
[ pylab.axvline(x*frmh, color="green") for x in edp ]
#pylab.xlabel("Czas [s]")
pylab.title("Energia wypowiedzi w czasie")
time = np.array(range(len(data)), dtype=float)/fs
sc_data = data - np.mean(data)
sc_data = sc_data/np.max(sc_data)
pylab.subplot(2,1,2)
pylab.plot(time, sc_data)
[ pylab.axvline(pos(x)/fs, color="red") for x in stp ]
[ pylab.axvline(pos(x)/fs, color="green") for x in edp ]
pylab.xlabel("Czas [s]")
pylab.title(u"Zapis sygnału")
pylab.show()
return [ data[pos(a):pos(b)] for a,b in zip(stp,edp) ]
def find_words(data, fs, num, **args):
min_silence_length = args["min_silence_length"] if "min_silence_length" in args else 5
_plot = args["plot"] if "plot" in args else False
_debug = False
frml = 0.03
frmh = 0.02
X = stft(data, fs, frml, frmh)
magnitude = np.absolute(X)
signal_energy = np.sum(magnitude, 1)
if "thresh" in args.keys():
thresh = args["thresh"]
else:
thresh = np.mean(signal_energy) - np.min(signal_energy)
en_th = (signal_energy > thresh).astype(np.int8)
rising_edge = np.diff(np.hstack(([0], en_th)))
falling_edge = np.diff(np.hstack((en_th, [0])))
(stp,) = np.nonzero(rising_edge > 0)
(edp,) = np.nonzero(falling_edge < 0)
if len(stp) != len(edp):
raise ValueError("Wrong threshold value")
"""Remove too short silence periods"""
silence = stp[1:] - edp[:-1]
(too_short,) = np.nonzero(silence <= min_silence_length)
stp = np.delete(stp, too_short + 1)
edp = np.delete(edp, too_short)
nonsilence = edp - stp
shortest = nonsilence.argsort()
if len(nonsilence) < num:
raise ValueError("You say less words than you ask me to find or you speak too fast")
""" Remove too short nonsilence """
stp = np.delete(stp, shortest[:-num])
edp = np.delete(edp, shortest[:-num])
pos = lambda a: fs * frmh * a
if _plot:
pylab.subplot(2, 1, 1)
etime = np.array(range(len(signal_energy)), dtype=float) * frmh
pylab.plot(etime, signal_energy / 1e6)
pylab.axhline(thresh / 1e6)
[pylab.axvline(x * frmh, color="red") for x in stp]
[pylab.axvline(x * frmh, color="green") for x in edp]
# pylab.xlabel("Czas [s]")
pylab.title("Energia wypowiedzi w czasie")
time = np.array(range(len(data)), dtype=float) / fs
sc_data = data - np.mean(data)
sc_data = sc_data / np.max(sc_data)
pylab.subplot(2, 1, 2)
pylab.plot(time, sc_data)
[pylab.axvline(pos(x) / fs, color="red") for x in stp]
[pylab.axvline(pos(x) / fs, color="green") for x in edp]
pylab.xlabel("Czas [s]")
pylab.title(u"Zapis sygnału")
pylab.show()
return [data[pos(a) : pos(b)] for a, b in zip(stp, edp)]
