def test_buf_size_not_power_of_two(self):
win_s = 320
hop_s = win_s // 2
try:
with self.assertRaises(RuntimeError):
pvoc(win_s, hop_s)
except AssertionError:
# when compiled with fftw3, aubio supports non power of two fft sizes
self.skipTest('creating aubio.pvoc with size %d did not fail' % win_s)
def test_buf_size_not_power_of_two(self):
win_s = 320
hop_s = win_s // 2
try:
with self.assertRaises(RuntimeError):
pvoc(win_s, hop_s)
except AssertionError:
# when compiled with fftw3, aubio supports non power of two fft sizes
self.skipTest('creating aubio.pvoc with size %d did not fail' % win_s)
def mfcc(self, **options):
nFilters = options.get("nFilters") or 40 # must be 40 for mfcc
nCoeffs = options.get("nCoefs") or 13
sourceBuffer = source(self.audioFilename, self.samplerate, self.hopSize)
pvocBuffer = pvoc(self.fftWindowSize, self.hopSize)
mfccBuffer = mfcc(self.fftWindowSize, nFilters, nCoeffs, self.samplerate)
mfccs = np.zeros([nCoeffs, ])
timings = []
frames = []
totalFrames = 0
while True:
samples, read = sourceBuffer()
spec = pvocBuffer(samples)
mfcc_out = mfccBuffer(spec)
mfccs = np.vstack((mfccs, mfcc_out))
totalFrames += read
timings += [float(totalFrames) / self.samplerate]
frames += [totalFrames]
if read < self.hopSize: break
return mfccs
def Extract_MFCC(self,player_pos_x,player_pos_y):
import sys
from aubio import source, pvoc, mfcc
from numpy import vstack, zeros, diff
player=[player_pos_x,player_pos_y]
target=[target_position_x,target_position_y]
distance=scipy.spatial.distance.euclidean(player, target)
n_filters = 40 # must be 40 for mfcc
n_coeffs = 13
source_filename = './Audios/Hello.wav'
samplerate = 44100
win_s = 512
hop_s = 128
s = source(source_filename, samplerate, hop_s)
samplerate = s.samplerate
p = pvoc(win_s, hop_s)
m = mfcc(win_s, n_filters, n_coeffs, samplerate)
mfccs = zeros([n_coeffs, ])
frames_read = 0
while True:
samples, read = s()
spec = p(samples)
mfcc_out = m(spec)
mfccs = vstack((mfccs, mfcc_out))
frames_read += read
if read < hop_s: break
if(distance==0):
factor=1
else:
factor=1/distance
return mfccs*factor
def vocoder(file_id):
in_file = '/tmp/don_robot/output/' + file_id + ".mp3"
out_file = '/tmp/don_robot/output/' + file_id + "-voc.mp3"
samplerate = 44100
f = source(in_file, samplerate, 256)
g = sink(out_file, samplerate)
total_frames, read = 0, 256
win_s = 512 # fft size
hop_s = win_s // 2 # hop size
pv = pvoc(win_s, hop_s) # phase vocoder
while read:
samples, read = f()
spectrum = pv(samples) # compute spectrum
spectrum.norm *= .5 # reduce amplitude a bit .8
spectrum.phas[:] = 0. # zero phase
new_samples = pv.rdo(spectrum) # compute modified samples
g(new_samples, read) # write to output
total_frames += read
format_str = "read {:d} samples from {:s}, written to {:s}"
print(format_str.format(total_frames, f.uri, g.uri))
return out_file
def get_spectrogram(row, col):
samplerate = 0
from aubio import pvoc, source, float_type
from numpy import zeros, log10, vstack
win_s = 512 # fft window size
hop_s = win_s // 2 # hop size
fft_s = win_s // 2 + 1 # spectrum bins
filename = 'Test.wav'
a = source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = pvoc(win_s, hop_s) # phase vocoder
specgram = zeros([0, fft_s],
dtype=float_type) # numpy array to store spectrogram
spec = []
# analysis
while True:
samples, read = a() # read file
specgram = vstack(
(specgram, pv(samples).norm)) # store new norm vector
if read < a.hop_size: break
if (row == grid_size.nRow - 1 and grid_size == grid_size.nCol - 1):
print("True")
print("row {} col {}".format(row, col))
specgram = specgram + 100
return specgram
def __init__(self):
self.redis = redis.StrictRedis(host=redishost, port=6379, password="", decode_responses=True)
self.p = pyaudio.PyAudio()
stream = self.p.open(format=self.FORMAT,
channels=self.CHANNELS,
rate=self.RATE,
input=True,
output=True,
input_device_index = self.get_input_device_index(),
output_device_index = self.get_output_device_index(),
frames_per_buffer = self.CHUNK,
stream_callback=self.callback)
self.a_onset = aubio.onset("default", self.CHUNK, self.hop_s, self.RATE)
self.a_tempo = aubio.tempo("specflux", self.CHUNK, self.hop_s, self.RATE)
self.a_pitch = aubio.pitch("default", self.CHUNK, self.hop_s, self.RATE)
self.a_notes = aubio.notes("default", self.CHUNK, self.hop_s, self.RATE)
n_filters = 40 # required
n_coeffs = 13 # I wonder if i made this 1....
self.a_pvoc = aubio.pvoc(self.CHUNK, self.hop_s)
self.a_mfcc = aubio.mfcc(self.CHUNK, n_filters, n_coeffs, self.RATE)
self.tolerance = 0.8
self.a_pitch.set_tolerance(self.tolerance)
self.highest_pitch = 0
self.lowest_pitch = 99999999
self.average_pitch = 0
self.average_pitch_samples = 0
self.last_average = 0
self.colors = None
self.pitch_range = None
self.range_counter = 0
self.all_notes = set()
stream.start_stream()
def Extract_MFCC(row, col):
import sys
from aubio import source, pvoc, mfcc
from numpy import vstack, zeros, diff
n_filters = 40 # must be 40 for mfcc
n_coeffs = 13
source_filename = 'Test.wav'
samplerate = 44100
win_s = 512
hop_s = 128
s = source(source_filename, samplerate, hop_s)
samplerate = s.samplerate
p = pvoc(win_s, hop_s)
m = mfcc(win_s, n_filters, n_coeffs, samplerate)
mfccs = zeros([
n_coeffs,
])
frames_read = 0
while True:
samples, read = s()
spec = p(samples)
mfcc_out = m(spec)
mfccs = vstack((mfccs, mfcc_out))
frames_read += read
if read < hop_s: break
if (row == grid_size.nRow - 1 and grid_size == grid_size.nCol - 1):
mfccs = mfccs + 100
return mfccs
def vocoder(file_id):
in_file = '/tmp/don_robot/output/' + file_id + ".mp3"
out_file = '/tmp/don_robot/output/' + file_id + "-voc.mp3"
samplerate = 44100
f = source(in_file, samplerate, 256)
g = sink(out_file, samplerate)
total_frames, read = 0, 256
win_s = 512 # fft size
hop_s = win_s // 2 # hop size
pv = pvoc(win_s, hop_s) # phase vocoder
while read:
samples, read = f()
spectrum = pv(samples) # compute spectrum
spectrum.norm *= .5 # reduce amplitude a bit .8
spectrum.phas[:] = 0. # zero phase
new_samples = pv.rdo(spectrum) # compute modified samples
g(new_samples, read) # write to output
total_frames += read
format_str = "read {:d} samples from {:s}, written to {:s}"
print(format_str.format(total_frames, f.uri, g.uri))
return out_file
def getMelEnergy(path):
win_s = 512 #fft size
hop_s = win_s / 4 #hop size
samplerate = 0
s = source(path, samplerate, hop_s)
samplerate = s.samplerate
pv = pvoc(win_s, hop_s)
f = filterbank(40, win_s)
f.set_mel_coeffs_slaney(samplerate)
energies = np.zeros((40, ))
o = {}
total_frames = 0
downsample = 2
while True:
samples, read = s()
fftgrain = pv(samples)
new_energies = f(fftgrain)
energies = np.vstack([energies, new_energies])
total_frames += read
if read < hop_s:
break
return energies
def aubio(source_filename):
# print("Usage: %s <source_filename> [samplerate] [win_s] [hop_s] [mode]" % sys.argv[0])
# print(" where [mode] can be 'delta' or 'ddelta' for first and second derivatives")7
n_filters = 40 # must be 40 for mfcc
n_coeffs = 26
win_s = 512
hop_s = win_s // 4
# mode = "default"
samplerate = 0
s = source(source_filename, samplerate, hop_s)
samplerate = s.samplerate
p = pvoc(win_s, hop_s)
m = mfcc(win_s, n_filters, n_coeffs, samplerate)
mfccs = zeros([
n_coeffs,
])
frames_read = 0
while True:
samples, read = s()
spec = p(samples)
mfcc_out = m(spec)
mfccs = vstack((mfccs, mfcc_out))
frames_read += read
if read < hop_s: break
return mfccs
def __init__(self,
samplerate=44100,
winsize=1024,
hopsize=512,
filters=40,
coeffs=13):
super(AubioAnalyser, self).__init__()
self.winsize = winsize
self.hopsize = hopsize
self.coeffs = coeffs
self.filters = filters
self.descriptors = {}
self.methods = [
"default", "energy", "hfc", "complex", "phase", "specdiff", "kl",
"mkl", "specflux", "centroid", "slope", "rolloff", "spread",
"skewness", "kurtosis", "decrease"
]
for method in self.methods:
self.descriptors[method] = aubio.specdesc(method, self.winsize)
self.pvocoder = aubio.pvoc(self.winsize, self.hopsize)
self.mfcc_feature = aubio.mfcc(winsize, filters, coeffs, samplerate)
self.mfccs = numpy.zeros([
self.coeffs,
])
def mfcc(frame, audiofile):
'''
Computes the MEL FREQUENCY CEPSTRAL COEFFICIENTS for the frame,
the frame is zero padded to achieve a frame lenght which is a power
of two if this is not already the case. The power spectrum is then computed
and this is placed into filterbanks on a mel-scale. The coefficents of
12 of the banks is then returned.
'''
coefficientsCount = 12
sampleRate = audiofile['sample_rate']
frame_size = audiofile['frame_size']
fftsize = pow(2, int(math.log(frame_size, 2) +
0.5)) # Round to nearest power of 2 to facilitate FFT
m = aub.mfcc(fftsize, 40, coefficientsCount, sampleRate)
#first we need to convert this frame to the power spectrum using a DFT
p = aub.pvoc(fftsize, int(frame_size))
#in order to compute DFT the frame must be of a length which is a power of 2, so expand to fftsize using zero padding
if len(frame) != 16000:
frame = np.pad(frame, (0, frame_size - len(frame)),
'constant',
constant_values=0)
#compute the power spectrum
spec = p(frame.astype(np.float32))
#compute the MFCC, which returns the coefficents of each of the 12 coefficents
mfcc_out = m(spec)
return mfcc_out
def dataIn():
pv = pvoc(self.win_s, self.hop_s)
while True:
samples, read = self.source() # Read the file
cvec = pv(samples)
if read < self.source.hop_size:
break # Out of samples
yield (cvec)
def test_no_overlap(self):
win_s, hop_s = 1024, 1024
f = pvoc(win_s, hop_s)
t = fvec(hop_s)
for _ in range(4):
s = f(t)
r = f.rdo(s)
assert_equal(t, 0.)
def test_no_overlap(self):
win_s, hop_s = 1024, 1024
f = pvoc (win_s, hop_s)
t = fvec (hop_s)
for _ in range(4):
s = f(t)
r = f.rdo(s)
assert_equal ( t, 0.)
def get_spectrogram(filename, samplerate = 0):
win_s = 512 # fft window size
hop_s = win_s // 2 # hop size
fft_s = win_s // 2 + 1 # spectrum bins
a = source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = pvoc(win_s, hop_s) # phase vocoder
specgram = zeros([0, fft_s], dtype=float_type) # numpy array to store spectrogram
# analysis
while True:
samples, read = a() # read file
specgram = vstack((specgram,pv(samples).norm)) # store new norm vector
if read < a.hop_size: break
# plotting
fig = plt.imshow(log10(specgram.T + .001), origin = 'bottom', aspect = 'auto', cmap=plt.cm.gray_r)
ax = fig.axes
ax.axis([0, len(specgram), 0, len(specgram[0])])
# show axes in Hz and seconds
time_step = hop_s / float(samplerate)
total_time = len(specgram) * time_step
outstr = "total time: %0.2fs" % total_time
print(outstr + ", samplerate: %.2fkHz" % (samplerate / 1000.))
n_xticks = 10
n_yticks = 10
def get_rounded_ticks( top_pos, step, n_ticks ):
top_label = top_pos * step
# get the first label
ticks_first_label = top_pos * step / n_ticks
# round to the closest .1
ticks_first_label = round ( ticks_first_label * 10. ) / 10.
# compute all labels from the first rounded one
ticks_labels = [ ticks_first_label * n for n in range(n_ticks) ] + [ top_label ]
# get the corresponding positions
ticks_positions = [ ticks_labels[n] / step for n in range(n_ticks) ] + [ top_pos ]
# convert to string
ticks_labels = [ "%.1f" % x for x in ticks_labels ]
# return position, label tuple to use with x/yticks
return ticks_positions, ticks_labels
# apply to the axis
x_ticks, x_labels = get_rounded_ticks ( len(specgram), time_step, n_xticks )
y_ticks, y_labels = get_rounded_ticks ( len(specgram[0]), (samplerate / 1000. / 2.) / len(specgram[0]), n_yticks )
ax.set_xticks( x_ticks )
ax.set_yticks ( y_ticks )
ax.set_xticklabels( x_labels )
ax.set_yticklabels ( y_labels )
ax.set_ylabel('Frequency (kHz)')
ax.set_xlabel('Time (s)')
ax.set_title(os.path.basename(filename))
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize('x-small')
return fig
def get_spectrogram(filename, samplerate=0):
win_s = 512 # fft window size
hop_s = win_s // 2 # hop size
fft_s = win_s // 2 + 1 # spectrum bins
a = source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = pvoc(win_s, hop_s) # phase vocoder
specgram = zeros([0, fft_s], dtype=float_type) # numpy array to store spectrogram
# analysis
while True:
samples, read = a() # read file
specgram = vstack((specgram, pv(samples).norm)) # store new norm vector
if read < a.hop_size: break
# plotting
fig = plt.imshow(log10(specgram.T + .001), origin='bottom', aspect='auto', cmap=plt.cm.gray_r)
ax = fig.axes
ax.axis([0, len(specgram), 0, len(specgram[0])])
# show axes in Hz and seconds
time_step = hop_s / float(samplerate)
total_time = len(specgram) * time_step
outstr = "total time: %0.2fs" % total_time
print(outstr + ", samplerate: %.2fkHz" % (samplerate / 1000.))
n_xticks = 10
n_yticks = 10
def get_rounded_ticks(top_pos, step, n_ticks):
top_label = top_pos * step
# get the first label
ticks_first_label = top_pos * step / n_ticks
# round to the closest .1
ticks_first_label = round(ticks_first_label * 10.) / 10.
# compute all labels from the first rounded one
ticks_labels = [ticks_first_label * n for n in range(n_ticks)] + [top_label]
# get the corresponding positions
ticks_positions = [ticks_labels[n] / step for n in range(n_ticks)] + [top_pos]
# convert to string
ticks_labels = ["%.1f" % x for x in ticks_labels]
# return position, label tuple to use with x/yticks
return ticks_positions, ticks_labels
# apply to the axis
x_ticks, x_labels = get_rounded_ticks(len(specgram), time_step, n_xticks)
y_ticks, y_labels = get_rounded_ticks(len(specgram[0]), (samplerate / 1000. / 2.) / len(specgram[0]), n_yticks)
ax.set_xticks(x_ticks)
ax.set_yticks(y_ticks)
ax.set_xticklabels(x_labels)
ax.set_yticklabels(y_labels)
ax.set_ylabel('Frequency (kHz)')
ax.set_xlabel('Time (s)')
ax.set_title(os.path.basename(filename))
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize('x-small')
return fig
def activate(self):
if self._audio is None:
self._audio = pyaudio.PyAudio()
# Setup a pre-emphasis filter to help balance the highs
self.pre_emphasis = None
if self._config['pre_emphasis']:
self.pre_emphasis = aubio.digital_filter(3)
# old, do not use
#self.pre_emphasis.set_biquad(1., -self._config['pre_emphasis'], 0, 0, 0)
# USE THESE FOR SCOTT_MEL OR OTHERS
#self.pre_emphasis.set_biquad(1.3662, -1.9256, 0.5621, -1.9256, 0.9283)
# USE THESE FOR MATT_MEl
# weaker bass, good for vocals, highs
#self.pre_emphasis.set_biquad(0.87492, -1.74984, 0.87492, -1.74799, 0.75169)
# bass heavier overall more balanced
self.pre_emphasis.set_biquad(0.85870, -1.71740, 0.85870, -1.71605,
0.71874)
# Setup the phase vocoder to perform a windowed FFT
self._phase_vocoder = aubio.pvoc(
self._config['fft_size'],
self._config['mic_rate'] // self._config['sample_rate'])
self._frequency_domain_null = aubio.cvec(self._config['fft_size'])
self._frequency_domain = self._frequency_domain_null
self._frequency_domain_x = np.linspace(
0, self._config['mic_rate'], (self._config["fft_size"] // 2) + 1)
# Enumerate all of the input devices and find the one matching the
# configured device index
_LOGGER.info("Audio Input Devices:")
info = self._audio.get_host_api_info_by_index(0)
for i in range(0, info.get('deviceCount')):
if (self._audio.get_device_info_by_host_api_device_index(
0, i).get('maxInputChannels')) > 0:
_LOGGER.info(" [{}] {}".format(
i,
self._audio.get_device_info_by_host_api_device_index(
0, i).get('name')))
# Open the audio stream and start processing the input
self._stream = self._audio.open(
input_device_index=self._config['device_index'],
format=pyaudio.paFloat32,
channels=1,
rate=self._config['mic_rate'],
input=True,
frames_per_buffer=self._config['mic_rate'] //
self._config['sample_rate'],
stream_callback=self._audio_sample_callback)
self._stream.start_stream()
_LOGGER.info("Audio source opened.")
def __init__(self, sample_rate=11025, buf_size=1024, hop_size=256,
n_mfcc_filters=40, n_mfcc_coeffs=13, pitch_method='yin'):
"""Initialize feature extractor."""
self._sample_rate = int(sample_rate)
self._pvoc = aubio.pvoc(buf_size, hop_size)
self._mfcc = aubio.mfcc(buf_size, n_mfcc_filters, n_mfcc_coeffs,
self._sample_rate)
self._pitch = aubio.pitch(
method=pitch_method, buf_size=buf_size, hop_size=hop_size,
samplerate=self._sample_rate)
def get_spectrogram(filename, samplerate=0):
win_s = 512 # fft window size
hop_s = win_s / 2 # hop size
fft_s = win_s / 2 + 1 # spectrum bins
a = source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = pvoc(win_s, hop_s) # phase vocoder
specgram = zeros([0, fft_s],
dtype='float32') # numpy array to store spectrogram
# analysis
while True:
samples, read = a() # read file
specgram = vstack(
(specgram, pv(samples).norm)) # store new norm vector
if read < a.hop_size: break
# plotting
imshow(log10(specgram.T + .001),
origin='bottom',
aspect='auto',
cmap=cm.gray_r)
axis([0, len(specgram), 0, len(specgram[0])])
# show axes in Hz and seconds
time_step = hop_s / float(samplerate)
total_time = len(specgram) * time_step
print "total time: %0.2fs" % total_time,
print ", samplerate: %.2fkHz" % (samplerate / 1000.)
n_xticks = 10
n_yticks = 10
def get_rounded_ticks(top_pos, step, n_ticks):
top_label = top_pos * step
# get the first label
ticks_first_label = top_pos * step / n_ticks
# round to the closest .1
ticks_first_label = round(ticks_first_label * 10.) / 10.
# compute all labels from the first rounded one
ticks_labels = [ticks_first_label * n
for n in range(n_ticks)] + [top_label]
# get the corresponding positions
ticks_positions = [ticks_labels[n] / step
for n in range(n_ticks)] + [top_pos]
# convert to string
ticks_labels = ["%.1f" % x for x in ticks_labels]
# return position, label tuple to use with x/yticks
return ticks_positions, ticks_labels
# apply to the axis
xticks(*get_rounded_ticks(len(specgram), time_step, n_xticks))
yticks(*get_rounded_ticks(len(specgram[0]), (samplerate / 2. / 1000.) /
len(specgram[0]), n_yticks))
ylabel('Frequency (kHz)')
xlabel('Time (s)')
def setup(self, channels=None, samplerate=None,
blocksize=None, totalframes=None):
super(AubioMelEnergy, self).setup(
channels, samplerate, blocksize, totalframes)
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.melenergy = filterbank(self.n_filters, self.input_blocksize)
self.melenergy.set_mel_coeffs_slaney(samplerate)
self.block_read = 0
self.melenergy_results = []
def activate(self):
if self._audio is None:
self._audio = pyaudio.PyAudio()
# Setup a pre-emphasis filter to help balance the highs
self.pre_emphasis = None
if self._config['pre_emphasis']:
self.pre_emphasis = aubio.digital_filter(3)
self.pre_emphasis.set_biquad(1., -self._config['pre_emphasis'], 0,
0, 0)
# Setup the phase vocoder to perform a windowed FFT
self._phase_vocoder = aubio.pvoc(
self._config['fft_size'],
self._config['mic_rate'] // self._config['sample_rate'])
self._frequency_domain_null = aubio.cvec(self._config['fft_size'])
self._frequency_domain = self._frequency_domain_null
self._frequency_domain_x = np.linspace(
0, self._config['mic_rate'], (self._config["fft_size"] // 2) + 1)
# Enumerate all of the input devices and find the one matching the
# configured device index
_LOGGER.info("Audio Input Devices:")
info = self._audio.get_host_api_info_by_index(0)
for i in range(0, info.get('deviceCount')):
if (self._audio.get_device_info_by_host_api_device_index(
0, i).get('maxInputChannels')) > 0:
_LOGGER.info(" [{}] {}".format(
i,
self._audio.get_device_info_by_host_api_device_index(
0, i).get('name')))
# PyAudio may segfault, reset device index if it seems implausible
if self._config['device_index'] >= info.get(
'deviceCount'
) or self._audio.get_device_info_by_host_api_device_index(
0, self._config['device_index']).get('maxInputChannels') <= 0:
_LOGGER.warn("Invalid device_index setting, resetting it to 0")
self._config['device_index'] = 0
# Open the audio stream and start processing the input
self._stream = self._audio.open(
input_device_index=self._config['device_index'],
format=pyaudio.paFloat32,
channels=1,
rate=self._config['mic_rate'],
input=True,
frames_per_buffer=self._config['mic_rate'] //
self._config['sample_rate'],
stream_callback=self._audio_sample_callback)
self._stream.start_stream()
_LOGGER.info("Audio source opened.")
def test_zeros(self):
win_s, hop_s = 1024, 256
f = pvoc(win_s, hop_s)
t = fvec(hop_s)
for time in range(4 * win_s / hop_s):
s = f(t)
r = f.rdo(s)
assert_equal(array(t), 0)
assert_equal(s.norm, 0)
assert_equal(s.phas, 0)
assert_equal(r, 0)
def setup(self, channels=None, samplerate=None,
blocksize=None, totalframes=None):
super(AubioMelEnergy, self).setup(
channels, samplerate, blocksize, totalframes)
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.melenergy = filterbank(self.n_filters, self.input_blocksize)
self.melenergy.set_mel_coeffs_slaney(samplerate)
self.block_read = 0
self.melenergy_results = []
def reconstruction(self, sigin, hop_s, ratio):
buf_s = hop_s * ratio
f = pvoc(buf_s, hop_s)
zeros = fvec(hop_s)
r2 = f.rdo(f(sigin))
for _ in range(1, ratio):
r2 = f.rdo(f(zeros))
# compute square errors
sq_error = (r2 - sigin)**2
# make sure all square errors are less than desired precision
assert_array_less(sq_error, max_sq_error)
def analyzeMFCC(grain):
windowSize = int(float(grain["frameCount"]))
s = source(grain["file"], int(grain["sampleRate"]), windowSize)
sampleRate = s.samplerate
p = pvoc(windowSize, windowSize)
m = mfcc(windowSize, 40, 13, s.samplerate)
samples, read = s()
spec = p(samples)
mfcc_out = m(spec)
mfccs = mfcc_out.tolist()
return mfccs
def reconstruction(self, sigin, hop_s, ratio):
buf_s = hop_s * ratio
f = pvoc(buf_s, hop_s)
zeros = fvec(hop_s)
r2 = f.rdo( f(sigin) )
for _ in range(1, ratio):
r2 = f.rdo( f(zeros) )
# compute square errors
sq_error = (r2 - sigin)**2
# make sure all square errors are less than desired precision
assert_array_less(sq_error, max_sq_error)
def analyzeMFCC(grain):
windowSize = int(float(grain["frameCount"]))
s = source(grain["file"], int(grain["sampleRate"]), windowSize - 1)
sampleRate = s.samplerate
p = pvoc(windowSize, windowSize - 1)
m = mfcc(windowSize, 40, 13, s.samplerate)
samples, read = s()
spec = p(samples)
mfcc_out = m(spec)
mfccs = mfcc_out.tolist()
return mfccs
def test_zeros(self):
win_s, hop_s = 1024, 256
f = pvoc (win_s, hop_s)
t = fvec (hop_s)
for time in range( 4 * win_s / hop_s ):
s = f(t)
r = f.rdo(s)
assert_equal ( array(t), 0)
assert_equal ( s.norm, 0)
assert_equal ( s.phas, 0)
assert_equal ( r, 0)
def __init__(self):
super(AubioMfcc, self).__init__()
self.input_blocksize = 1024
self.input_stepsize = self.input_blocksize // 4
# Aubio MFCC Initialisation
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.mfcc = None
self.block_read = 0
self.mfcc_results = np.zeros([self.n_coeffs, ])
def test_steps_three_random_channels(self):
from random import random
f = pvoc(64, 16)
t0 = fvec(16)
t1 = fvec(16)
for i in xrange(16):
t1[i] = random() * 2. - 1.
t2 = f.rdo(f(t1))
t2 = f.rdo(f(t0))
t2 = f.rdo(f(t0))
t2 = f.rdo(f(t0))
assert_almost_equal(t1, t2, decimal=6)
def test_zeros(self):
""" check the resynthesis of zeros gives zeros """
win_s, hop_s = 1024, 256
f = pvoc(win_s, hop_s)
t = fvec(hop_s)
for time in range(4 * win_s / hop_s):
s = f(t)
r = f.rdo(s)
assert_equal(array(t), 0)
assert_equal(s.norm, 0)
assert_equal(s.phas, 0)
assert_equal(r, 0)
def __init__(self):
super(AubioMfcc, self).__init__()
self.input_blocksize = 1024
self.input_stepsize = self.input_blocksize / 4
# Aubio MFCC Initialisation
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.mfcc = None
self.block_read = 0
self.mfcc_results = np.zeros([self.n_coeffs, ])
def test_zeros(self):
""" check the resynthesis of zeros gives zeros """
win_s, hop_s = 1024, 256
f = pvoc (win_s, hop_s)
t = fvec (hop_s)
for time in range( 4 * win_s / hop_s ):
s = f(t)
r = f.rdo(s)
assert_equal ( array(t), 0)
assert_equal ( s.norm, 0)
assert_equal ( s.phas, 0)
assert_equal ( r, 0)
def test_steps_three_random_channels(self):
from random import random
f = pvoc(64, 16)
t0 = fvec(16)
t1 = fvec(16)
for i in xrange(16):
t1[i] = random() * 2. - 1.
t2 = f.rdo(f(t1))
t2 = f.rdo(f(t0))
t2 = f.rdo(f(t0))
t2 = f.rdo(f(t0))
assert_almost_equal( t1, t2, decimal = 6 )
def __init__(self):
super(AubioMelEnergy, self).__init__()
self.input_blocksize = 1024
self.input_stepsize = self.input_blocksize / 4
# Aubio Melenergy Initialisation
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.melenergy = filterbank(self.n_filters, self.input_blocksize)
self.block_read = 0
self.melenergy_results = []
def __init__(self):
super(AubioMelEnergy, self).__init__()
self.input_blocksize = 1024
self.input_stepsize = self.input_blocksize / 4
# Aubio Melenergy Initialisation
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.melenergy = filterbank(self.n_filters, self.input_blocksize)
self.block_read = 0
self.melenergy_results = []
def setup(self, channels=None, samplerate=None,
blocksize=None, totalframes=None):
super(AubioMfcc, self).setup(
channels, samplerate, blocksize, totalframes)
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.mfcc = mfcc(self.input_blocksize,
self.n_filters,
self.n_coeffs,
samplerate)
self.block_read = 0
self.mfcc_results = numpy.zeros([self.n_coeffs, ])
def setup(self, channels=None, samplerate=None,
blocksize=None, totalframes=None):
super(AubioMfcc, self).setup(
channels, samplerate, blocksize, totalframes)
self.n_filters = 40
self.n_coeffs = 13
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.mfcc = mfcc(self.input_blocksize,
self.n_filters,
self.n_coeffs,
samplerate)
self.block_read = 0
self.mfcc_results = numpy.zeros([self.n_coeffs, ])
def mfccs(windows):
p = aubio.pvoc(window_size, hop_size)
m = aubio.mfcc(window_size, n_filters, n_coeffs, sample_rate)
windows = np.float32(windows)
mfccs = np.zeros([13,])
for i in range(len(windows)):
samples = windows[i]
spec = p(samples)
mfcc_out = m(spec)
mfccs = np.vstack((mfccs, mfcc_out))
del p
del m
return mfccs
def __init__(self, args):
valid_opts1 = ['hop_size', 'buf_size']
self.parse_options(args, valid_opts1)
self.remap_pvoc_options(self.options)
self.pv = aubio.pvoc(**self.options)
valid_opts2 = ['buf_size', 'n_filters', 'n_coeffs', 'samplerate']
self.parse_options(args, valid_opts2)
self.mfcc = aubio.mfcc(**self.options)
# remember all options
self.parse_options(args, list(set(valid_opts1 + valid_opts2)))
super(process_mfcc, self).__init__(args)
def test_resynth_three_steps(self):
""" check the resynthesis of steps is correct with 25% overlap """
hop_s = 16
buf_s = hop_s * 4
sigin = fvec(hop_s)
zeros = fvec(hop_s)
f = pvoc(buf_s, hop_s)
for i in xrange(hop_s):
sigin[i] = random() * 2. - 1.
t2 = f.rdo( f(sigin) )
t2 = f.rdo( f(zeros) )
t2 = f.rdo( f(zeros) )
t2 = f.rdo( f(zeros) )
assert_almost_equal( sigin, t2, decimal = precision )
def __init__(self, args):
self.args = args
valid_opts = ['hop_size', 'buf_size']
self.parse_options(args, valid_opts)
self.remap_pvoc_options(self.options)
self.pv = aubio.pvoc(**self.options)
valid_opts = ['buf_size', 'n_filters']
self.parse_options(args, valid_opts)
self.remap_pvoc_options(self.options)
self.filterbank = aubio.filterbank(**self.options)
self.filterbank.set_mel_coeffs_slaney(args.samplerate)
super(process_melbands, self).__init__(args)
def __init__(self, args):
valid_opts1 = ['hop_size', 'buf_size']
self.parse_options(args, valid_opts1)
self.remap_pvoc_options(self.options)
self.pv = aubio.pvoc(**self.options)
valid_opts2 = ['buf_size', 'n_filters', 'n_coeffs', 'samplerate']
self.parse_options(args, valid_opts2)
self.mfcc = aubio.mfcc(**self.options)
# remember all options
self.parse_options(args, list(set(valid_opts1 + valid_opts2)))
super(process_mfcc, self).__init__(args)
def __init__(self, args):
self.args = args
valid_opts = ['hop_size', 'buf_size']
self.parse_options(args, valid_opts)
self.remap_pvoc_options(self.options)
self.pv = aubio.pvoc(**self.options)
valid_opts = ['buf_size', 'n_filters']
self.parse_options(args, valid_opts)
self.remap_pvoc_options(self.options)
self.filterbank = aubio.filterbank(**self.options)
self.filterbank.set_mel_coeffs_slaney(args.samplerate)
super(process_melbands, self).__init__(args)
def test_resynth_three_steps(self):
""" check the resynthesis of steps is correct with 25% overlap """
hop_s = 16
buf_s = hop_s * 4
sigin = fvec(hop_s)
zeros = fvec(hop_s)
f = pvoc(buf_s, hop_s)
for i in range(hop_s):
sigin[i] = random() * 2. - 1.
t2 = f.rdo(f(sigin))
t2 = f.rdo(f(zeros))
t2 = f.rdo(f(zeros))
t2 = f.rdo(f(zeros))
assert_almost_equal(sigin, t2, decimal=precision)
def __init__(self, options):
if options.settings['fbuckets'] == 'third-octave':
options.settings['fbuckets'] = [22.4,
25, 31.5, 40, 50, 63,
80, 100, 125, 160, 200,
250, 315, 400, 500, 630,
800, 1000, 1250, 1600, 2000,
2500, 3150, 4000, 5000, 6300,
8000, 10000, 12500, 16000, 20000,
22390]
elif options.settings['fbuckets'] == 'octave':
options.settings['fbuckets'] = [22,
31.5, 63, 125, 250, 500,
1000, 2000, 4000, 8000, 16000,
22720]
self.midi_processor = None
self.sysex_command_array = []
for command in options.settings['fsysexnum'].split(' '):
self.sysex_command_array.append(int(command, 0))
self.filter_bank = filterbank(len(options.settings['fbuckets']) - 2,
(int(options.settings['framesize']) *
int(options.settings['fframemult'])))
self.frequencies = fvec(options.settings['fbuckets'])
self.filter_bank.set_triangle_bands(self.frequencies,
int(options.settings[
'samplerate']))
self.phase_vocoder = pvoc(int(float(options.settings['framesize']) *
float(options.settings['fframemult'])),
int(float(options.settings['framesize']) *
float(options.settings['fframemult']) *
float(options.settings['fhopmult'])))
self.frame_arrays = np.zeros(
(int(float(options.settings['fframemult'])),
int(float(options.settings['framesize']))),
dtype=np.float32)
self.frame_count = 0
self.maximum_frequencies = np.zeros(
(len(options.settings['fbuckets']) - 2,), dtype=np.float32)
self.last_energies = np.zeros((len(options.settings['fbuckets']) - 2,),
dtype=np.float32)
self.count_energies = np.zeros((int(options.settings['fcount']),
(len(options.settings[
'fbuckets']) - 2)),
dtype=np.float32)
self.energy_count = 0
self.rest_stop = 0
self.frame_multiplier = int(options.settings['fframemult'])
self.count = int(options.settings['fcount'])
self.graceful = float(options.settings['fgraceful'])
def __iter__(self):
if self.verbose:
print('[DEBUG] PhaseVocPR.__iter__()')
pv = pvoc(self.win_s, self.hop_s)
if self.verbose:
print(
' Created Phase Vocoder (pv = pvoc(self.win_s, self.hop_s)'
)
while True:
samples, read = self.source() # Read the file
cvec = pv(samples)
if read < self.source.hop_size:
break # Out of samples
yield (cvec)
def test_steps_two_channels(self): """ check the resynthesis of steps is correct """ f = pvoc(1024, 512) t1 = fvec(512) t2 = fvec(512) # positive step in first channel t1[100:200] = .1 # positive step in second channel t1[20:50] = -.1 s1 = f(t1) r1 = f.rdo(s1) s2 = f(t2) r2 = f.rdo(s2) #self.plot_this ( s1.norm.T ) assert_almost_equal ( t1, r2, decimal = 6 )
def test_steps_two_channels(self):
""" check the resynthesis of steps is correct """
f = pvoc(1024, 512)
t1 = fvec(512)
t2 = fvec(512)
# positive step in first channel
t1[100:200] = .1
# positive step in second channel
t1[20:50] = -.1
s1 = f(t1)
r1 = f.rdo(s1)
s2 = f(t2)
r2 = f.rdo(s2)
#self.plot_this ( s1.norm.T )
assert_almost_equal(t1, r2, decimal=6)
def mfccs(windows):
p = aubio.pvoc(window_size, hop_size)
m = aubio.mfcc(window_size, n_filters, n_coeffs, sample_rate)
windows = np.float32(windows)
mfccs = np.zeros([
13,
])
for i in range(len(windows)):
samples = windows[i]
spec = p(samples)
mfcc_out = m(spec)
mfccs = np.vstack((mfccs, mfcc_out))
del p
del m
return mfccs
def get_spectrogram(filename, samplerate = 0):
win_s = 512 # fft window size
hop_s = win_s / 2 # hop size
fft_s = win_s / 2 + 1 # spectrum bins
a = source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = pvoc(win_s, hop_s) # phase vocoder
specgram = zeros([0, fft_s], dtype='float32') # numpy array to store spectrogram
# analysis
while True:
samples, read = a() # read file
specgram = vstack((specgram,pv(samples).norm)) # store new norm vector
if read < a.hop_size: break
# plotting
imshow(log10(specgram.T + .001), origin = 'bottom', aspect = 'auto', cmap=cm.gray_r)
axis([0, len(specgram), 0, len(specgram[0])])
# show axes in Hz and seconds
time_step = hop_s / float(samplerate)
total_time = len(specgram) * time_step
print "total time: %0.2fs" % total_time,
print ", samplerate: %.2fkHz" % (samplerate / 1000.)
n_xticks = 10
n_yticks = 10
def get_rounded_ticks( top_pos, step, n_ticks ):
top_label = top_pos * step
# get the first label
ticks_first_label = top_pos * step / n_ticks
# round to the closest .1
ticks_first_label = round ( ticks_first_label * 10. ) / 10.
# compute all labels from the first rounded one
ticks_labels = [ ticks_first_label * n for n in range(n_ticks) ] + [ top_label ]
# get the corresponding positions
ticks_positions = [ ticks_labels[n] / step for n in range(n_ticks) ] + [ top_pos ]
# convert to string
ticks_labels = [ "%.1f" % x for x in ticks_labels ]
# return position, label tuple to use with x/yticks
return ticks_positions, ticks_labels
# apply to the axis
xticks( *get_rounded_ticks ( len(specgram), time_step, n_xticks ) )
yticks( *get_rounded_ticks ( len(specgram[0]), (samplerate / 2. / 1000.) / len(specgram[0]), n_yticks ) )
ylabel('Frequency (kHz)')
xlabel('Time (s)')
def test_resynth_two_steps(self):
""" check the resynthesis of steps is correct with 50% overlap """
hop_s = 512
buf_s = hop_s * 2
f = pvoc(buf_s, hop_s)
sigin = fvec(hop_s)
zeros = fvec(hop_s)
# negative step
sigin[20:50] = -.1
# positive step
sigin[100:200] = .1
s1 = f(sigin)
r1 = f.rdo(s1)
s2 = f(zeros)
r2 = f.rdo(s2)
#self.plot_this ( s2.norm.T )
assert_almost_equal ( r2, sigin, decimal = precision )
def __init__(self):
super(AubioSpecdesc, self).__init__()
self.input_blocksize = 1024
self.input_stepsize = self.input_blocksize / 4
# Aubio Specdesc Initialisation
self.block_read = 0
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.methods = [
'default', 'energy', 'hfc', 'complex', 'phase', 'specdiff', 'kl',
'mkl', 'specflux', 'centroid', 'slope', 'rolloff', 'spread', 'skewness',
'kurtosis', 'decrease']
self.specdesc = {}
self.specdesc_results = {}
for method in self.methods:
self.specdesc[method] = specdesc(method, self.input_blocksize)
self.specdesc_results[method] = []
def get_spectrogram(filename, samplerate=0):
win_s = 320 # fft window size
hop_s = 160 # hop size
fft_s = win_s // 2 + 1 # spectrum bins
a = aubio.source(filename, samplerate, hop_s) # source file
if samplerate == 0: samplerate = a.samplerate
pv = aubio.pvoc(win_s, hop_s) # phase vocoder
specgram = np.zeros([0, fft_s], dtype=aubio.float_type) # numpy array to store spectrogram
# analysis
while True:
samples, read = a() # read file
sp0 = pv(samples)
specgram = np.vstack((specgram, sp0.norm)) # store new norm vector
if read < a.hop_size: break
np.savetxt("specgram.txt", specgram.flatten(), fmt='%f')
def __init__(self, samplerate=44100, winsize=1024, hopsize=512, filters=40,
coeffs=13):
super(AubioAnalyser, self).__init__()
self.winsize = winsize
self.hopsize = hopsize
self.coeffs = coeffs
self.filters = filters
self.descriptors = {}
self.methods = ["default", "energy", "hfc", "complex", "phase",
"specdiff", "kl", "mkl", "specflux", "centroid",
"slope", "rolloff", "spread", "skewness", "kurtosis",
"decrease"]
for method in self.methods:
self.descriptors[method] = aubio.specdesc(method, self.winsize)
self.pvocoder = aubio.pvoc(self.winsize, self.hopsize)
self.mfcc_feature = aubio.mfcc(winsize, filters, coeffs, samplerate)
self.mfccs = numpy.zeros([self.coeffs, ])
def test_zeros(self):
""" check the resynthesis of zeros gives zeros """
win_s, hop_s = 1024, 256
f = pvoc (win_s, hop_s)
t = fvec (hop_s)
for _ in range( int ( 4 * win_s / hop_s ) ):
s = f(t)
r = f.rdo(s)
assert_equal ( t, 0.)
assert_equal ( s.norm, 0.)
try:
assert_equal ( s.phas, 0 )
except AssertionError:
assert_equal (s.phas[s.phas > 0], +np.pi)
assert_equal (s.phas[s.phas < 0], -np.pi)
assert_equal (np.abs(s.phas[np.abs(s.phas) != np.pi]), 0)
self.skipTest('pvoc(fvec(%d)).phas != +0, ' % win_s \
+ 'This is expected when using fftw3 on powerpc.')
assert_equal ( r, 0.)
def setup(self, channels=None, samplerate=None,
blocksize=None, totalframes=None):
super(
AubioSpecdesc,
self).setup(
channels,
samplerate,
blocksize,
totalframes)
self.block_read = 0
self.pvoc = pvoc(self.input_blocksize, self.input_stepsize)
self.methods = [
'default', 'energy', 'hfc', 'complex', 'phase', 'specdiff', 'kl',
'mkl', 'specflux', 'centroid', 'slope', 'rolloff', 'spread', 'skewness',
'kurtosis', 'decrease']
self.specdesc = {}
self.specdesc_results = {}
for method in self.methods:
self.specdesc[method] = specdesc(method, self.input_blocksize)
self.specdesc_results[method] = []
def __init__(self, sample_rate=None, buffersize=None):
self.sample_rate = sample_rate or self.sample_rate
self.buffersize = buffersize or self.buffersize
self.window_size = self.buffersize * 2
self.stream = None
self.onset = aubio.onset(
'specflux', self.window_size, self.buffersize, self.sample_rate)
self.onset.set_threshold(0.3)
self.onset.set_silence(-20.)
self.tempo = aubio.tempo(
'default', self.window_size, self.buffersize, self.sample_rate)
self.energy = aubio.specdesc('specflux', self.buffersize * 2)
self.pv = aubio.pvoc(self.buffersize * 2, self.buffersize)
self.pitch = aubio.pitch(
"yinfft", self.window_size, self.buffersize, self.sample_rate)
self.pitch.set_unit("midi")
self.pitch.set_tolerance(0.8)
self.py_audio = pyaudio.PyAudio()