def note_specgram(path, ax, peak=70.0, use_cqt=True):
# Add several samples together
if isinstance(path, list):
for i, p in enumerate(path):
sr, a = readwav(f)
audio = a if i == 0 else a + audio
# Load one sample
else:
sr, audio = readwav(f)
audio = audio.astype(np.float32)
if use_cqt:
C = librosa.cqt(audio,
sr=sr,
hop_length=hop_length,
bins_per_octave=int(notes_per_octave * over_sample),
n_bins=int(octaves * notes_per_octave * over_sample),
real=False,
filter_scale=res_factor,
fmin=librosa.note_to_hz('C2'))
else:
C = librosa.stft(audio,
n_fft=n_fft,
win_length=n_fft,
hop_length=hop_length,
center=True)
mag, phase = librosa.core.magphase(C)
phase_angle = np.angle(phase)
phase_unwrapped = np.unwrap(phase_angle)
dphase = phase_unwrapped[:, 1:] - phase_unwrapped[:, :-1]
dphase = np.concatenate([phase_unwrapped[:, 0:1], dphase], axis=1) / np.pi
mag = (librosa.logamplitude(
mag**2, amin=1e-13, top_db=peak, ref_power=np.max) / peak) + 1
ax.matshow(dphase[::-1, :], cmap=plt.cm.rainbow)
ax.matshow(mag[::-1, :], cmap=my_mask)
def read_audio(path):
"""
:param path: Can be a list, or a single string
:return: sr, audio
"""
# Add several samples together
if isinstance(path, list):
for i, p in enumerate(path):
sr, a = readwav(path)
audio = a if i == 0 else a + audio
# Load one sample
else:
sr, audio = readwav(path)
audio = audio.astype(np.float32)
return sr, audio
def make_step(self, size_in, size_out, dt, rng):
assert size_in[0] == 0
assert size_out[0] == 1
rate = 1. / dt
orig_rate, orig = readwav(self.path)
new_size = int(orig.size * (rate / orig_rate))
wave = resample(orig, new_size)
wave -= wave.mean()
# Normalize wave to desired rms
wave_rms = npext.rms(wave)
wave *= (self.rms / wave_rms)
if self.at_end == 'loop':
def step_wavfileloop(t):
idx = int(t * rate) % wave.size
return wave[idx]
return step_wavfileloop
elif self.at_end == 'stop':
def step_wavfilestop(t):
idx = int(t * rate)
if idx > wave.size:
return 0.
else:
return wave[idx]
return step_wavfilestop
def make_step(self, size_in, size_out, dt, rng):
assert size_in[0] == 0
assert size_out[0] == 1
rate = 1. / dt
orig_rate, orig = readwav(self.path)
new_size = int(orig.size * (rate / orig_rate))
wave = resample(orig, new_size)
wave -= wave.mean()
# Normalize wave to desired rms
wave_rms = npext.rms(wave)
wave *= (self.rms / wave_rms)
if self.at_end == 'loop':
def step_wavfileloop(t):
idx = int(t * rate) % wave.size
return wave[idx]
return step_wavfileloop
elif self.at_end == 'stop':
def step_wavfilestop(t):
idx = int(t * rate)
if idx > wave.size:
return 0.
else:
return wave[idx]
return step_wavfilestop
def main():
import argparse
from scipy.io.wavfile import read as readwav
import numpy as np
parser = argparse.ArgumentParser(description='Compute the SRMR metric for a given WAV file')
parser.add_argument('-f', '--fast', dest='fast', action='store_true', default=False,
help='Use the faster version based on the gammatonegram')
parser.add_argument('-n', '--norm', dest='norm', action='store_true', default=False,
help='Use modulation spectrum energy normalization')
parser.add_argument('--ncochlearfilters', dest='n_cochlear_filters', type=int, default=23,
help='Number of filters in the acoustic filterbank')
parser.add_argument('--mincf', dest='min_cf', type=float, default=4.0,
help='Center frequency of the first modulation filter')
parser.add_argument('--maxcf', dest='max_cf', type=float, default=128.0,
help='Center frequency of the last modulation filter')
parser.add_argument('path', metavar='path', nargs='+',
help='Path of the file or files to be processed. Can also be a folder.')
args = parser.parse_args()
for f in args.path:
fs, s = readwav(f)
if np.issubdtype(s.dtype, np.int):
s = s.astype('float')/np.iinfo(s.dtype).max
srmr = SRMR(fs,
n_cochlear_filters=args.n_cochlear_filters,
min_cf=args.min_cf,
max_cf=args.max_cf,
fast=args.fast,
norm=args.norm)
out = srmr.predict(s, s, s)
r, energy = out['p']['srmr'], out['avg_energy']
print('%s, %f' % (f, r))
def load_wavs_from_dir(path):
fnames = map(lambda f: f.path,
filter(lambda f: f.is_file(), os.scandir(path)))
fnames, rates, wavs = map(
np.asarray, zip(*tuple(map(lambda f: (f, *readwav(f)), fnames))))
wrong_rates = (rates != sample_rate)
if wrong_rates.any():
print('Following files have wrong simple rate (!= {}):'.format(
sample_rate) + '\n\t' + '\n\t'.join(
map(lambda t: '"{}" | rate = {}'.format(*t),
np.vstack((fnames, rates))[:, wrong_rates].transpose())),
file=stderr)
print('\nSkipping them')
wavs = wavs[~wrong_rates]
maxlen = max(i.shape[0] for i in wavs)
ar = []
for i in wavs:
s = i[:] if len(i.shape) == 1 else i[:, 0] # Use single channel
s = s.copy()
s.resize((maxlen, ))
ar.append(s)
data = np.vstack(ar)
return data
def readwav(filename):
"""Read a WAV file and returns the data and sample rate
::
from spectrum.io import readwav
readwav()
"""
from scipy.io.wavfile import read as readwav
samplerate, signal = readwav(filename)
return signal, samplerate
def process_file(f, args):
fs, s = readwav(f)
if len(s.shape) > 1:
s = s[:,0]
if np.issubdtype(s.dtype, np.int):
s = s.astype('float')/np.iinfo(s.dtype).max
r, energy = srmr(s, fs, n_cochlear_filters=args.n_cochlear_filters,
min_cf=args.min_cf,
max_cf=args.max_cf,
fast=args.fast,
norm=args.norm)
return f, r
def process_file(f, args):
fs, s = readwav(f)
if len(s.shape) > 1:
s = s[:,0]
if np.issubdtype(s.dtype, np.int):
s = s.astype('float')/np.iinfo(s.dtype).max
r, energy = srmr(s, fs, n_cochlear_filters=args.n_cochlear_filters,
min_cf=args.min_cf,
max_cf=args.max_cf,
fast=args.fast,
norm=args.norm)
return f, r
def cqt(fname, dst):
sr, audio = readwav(fname)
CQT = librosa.amplitude_to_db(np.abs(
librosa.cqt(audio.astype(np.float32), sr=sr)),
ref=np.max)
fig, ax = plt.subplots(figsize=(6, 6))
d = librosa.display.specshow(CQT, y_axis='cqt_note', ax=ax)
fig.colorbar(d, ax=ax, format='%+2.0f dB')
ax.set_title('Constant-Q power spectrogram (note)')
out_path = dst / os.path.basename(fname).replace('.wav', '_cqt.jpg')
plt.savefig(out_path, dpi=100)
def chroma(fname, dst):
sr, audio = readwav(fname)
C = librosa.feature.chroma_cqt(y=audio.astype(np.float32), sr=sr)
fig, ax = plt.subplots(figsize=(6, 6))
d = librosa.display.specshow(C, y_axis='chroma')
fig.colorbar(d, ax=ax)
ax.set_title('Chromagram')
if not os.path.exists(dst):
os.makedirs(dst)
out_path = dst / os.path.basename(fname).replace('.wav', '_chroma.jpg')
plt.savefig(out_path, dpi=100)
def rainbow(fname, dst, peak=70.0, use_cqt=True):
# Constants
n_fft = 512
hop_length = 256
over_sample = 4
res_factor = 0.8
octaves = 6
notes_per_octave = 10
# Plotting functions
cdict = {
'red': ((0.0, 0.0, 0.0), (1.0, 0.0, 0.0)),
'green': ((0.0, 0.0, 0.0), (1.0, 0.0, 0.0)),
'blue': ((0.0, 0.0, 0.0), (1.0, 0.0, 0.0)),
'alpha': ((0.0, 1.0, 1.0), (1.0, 0.0, 0.0))
}
my_mask = matplotlib.colors.LinearSegmentedColormap('MyMask', cdict)
plt.register_cmap(cmap=my_mask)
fig, ax = plt.subplots(figsize=(6, 6))
sr, audio = readwav(fname)
audio = audio.astype(np.float32)
if use_cqt:
C = librosa.cqt(audio,
sr=sr,
hop_length=hop_length,
bins_per_octave=int(notes_per_octave * over_sample),
n_bins=int(octaves * notes_per_octave * over_sample),
filter_scale=res_factor,
fmin=librosa.note_to_hz('C2'))
else:
C = librosa.stft(audio,
n_fft=n_fft,
win_length=n_fft,
hop_length=hop_length,
center=True)
mag, phase = librosa.core.magphase(C)
phase_angle = np.angle(phase)
phase_unwrapped = np.unwrap(phase_angle)
dphase = phase_unwrapped[:, 1:] - phase_unwrapped[:, :-1]
dphase = np.concatenate([phase_unwrapped[:, 0:1], dphase], axis=1) / np.pi
mag = (librosa.power_to_db(mag**2, amin=1e-13, top_db=peak, ref=np.max) /
peak) + 1
ax.imshow(np.flipud(dphase[::-1, :]), cmap=plt.cm.rainbow, origin='lower')
ax.imshow(np.flipud(mag[::-1, :]), cmap=my_mask, origin='lower')
ax.set_title(f"Rainbow Spectrogram for {os.path.basename(fname)}")
out_path = dst / os.path.basename(fname).replace('.wav', '_rainbow.jpg')
plt.savefig(out_path, dpi=100)
def note_specgram(path, ax, use_cqt=True):
# Add several samples together
if isinstance(path, list):
for i, f in enumerate(path):
sr, a = readwav(f)
audio = a if i == 0 else a + audio
# Load one sample
else:
sr, audio = readwav(path)
audio = audio.astype(np.float32)
# Constants
hop_length = 256
if use_cqt:
over_sample = 4
res_factor = 1.0 # 0.8
octaves = 6
notes_per_octave = 12
C = librosa.cqt(audio,
sr=sr,
hop_length=hop_length,
bins_per_octave=int(notes_per_octave * over_sample),
n_bins=int(octaves * notes_per_octave * over_sample),
filter_scale=res_factor,
fmin=librosa.note_to_hz('C3'))
else:
n_fft = 512
C = librosa.stft(audio,
n_fft=n_fft,
win_length=n_fft,
hop_length=hop_length,
center=True)
plot_rainbow(ax, C)
# Split signal in frames
framelen = int(framelen * fs)
frames = segment_axis(x, length=framelen, overlap=0, end='pad')
frames_zero_mean = frames - frames.mean(axis=0)
frame_energy = 10 * np.log10(1 / (framelen - 1) *
(frames_zero_mean**2).sum(axis=1) + 1e-6)
max_energy = max(frame_energy)
speech_presence = (frame_energy > max_energy - theta_main) & (frame_energy
> theta_min)
x_vad = np.zeros_like(x, dtype=bool)
for idx, frame in enumerate(frames):
if speech_presence[idx]:
x_vad[idx * framelen:(idx + 1) * framelen] = True
else:
x_vad[idx * framelen:(idx + 1) * framelen] = False
return x[x_vad], x_vad
if __name__ == '__main__':
import sys
from scipy.io.wavfile import read as readwav
from matplotlib import pyplot as plt
fs, s = readwav(sys.argv[1])
s = s.astype('float') / np.iinfo(s.dtype).max
s_vad, speech_presence = simple_energy_vad(s, fs)
plt.plot(s)
plt.plot(s_vad - 1, 'g')
plt.show()
this is the benchmark method, not the method proposed by the authors).
'''
# Split signal in frames
framelen = int(framelen * fs)
frames = segment_axis(x, length=framelen, overlap=0, end='pad')
frames_zero_mean = frames - frames.mean(axis=0)
frame_energy = 10*np.log10(1/(framelen-1) * (frames_zero_mean**2).sum(axis=1) + 1e-6)
max_energy = max(frame_energy)
speech_presence = (frame_energy > max_energy - theta_main) & (frame_energy > theta_min)
x_vad = np.zeros_like(x, dtype=bool)
for idx, frame in enumerate(frames):
if speech_presence[idx]:
x_vad[idx*framelen:(idx+1)*framelen] = True
else:
x_vad[idx*framelen:(idx+1)*framelen] = False
return x[x_vad], x_vad
if __name__ == '__main__':
import sys
from scipy.io.wavfile import read as readwav
from matplotlib import pyplot as plt
fs, s = readwav(sys.argv[1])
s = s.astype('float')/np.iinfo(s.dtype).max
s_vad, speech_presence = simple_energy_vad(s, fs)
plt.plot(s)
plt.plot(s_vad - 1, 'g')
plt.show()