def test_stft_istft(self):
try:
import librosa
ds = F.load_digit_wav()
name = ds.keys()[0]
path = ds[name]
y, _ = speech.read(path, pcm=True)
hop_length = int(0.01 * 8000)
stft = signal.stft(y, n_fft=256, hop_length=hop_length, window='hann')
stft_ = librosa.stft(y, n_fft=256, hop_length=hop_length, window='hann')
self.assertTrue(np.allclose(stft, stft_.T))
y1 = signal.istft(stft, hop_length=hop_length, window='hann')
y2 = librosa.istft(stft_, hop_length=hop_length, window='hann')
self.assertTrue(np.allclose(y1, y2))
except ImportError:
print("test_stft_istft require librosa.")
def map(self, job):
'''
Return
------
[(name, spec, mspec, mfcc, pitch, vad), ...]
'''
audio_path, segments = job[0] if len(job) == 1 else job
try:
# load audio data
s, sr_orig = speech.read(audio_path)
if sr_orig is not None and self.sr is not None and \
sr_orig != self.sr:
raise Exception('Given sample rate (%d Hz) is different from '
'audio file sample rate (%d Hz).' %
(self.sr, sr_orig))
if sr_orig is None:
sr_orig = self.sr
N = len(s)
# processing all segments
ret = []
for name, start, end, channel in segments:
start = int(float(start) * sr_orig)
end = int(N if end <= 0 else float(end) * sr_orig)
data = s[start:end, channel] if s.ndim > 1 else s[start:end]
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
features = speech.speech_features(
data.ravel(),
sr=sr_orig,
win=self.win,
shift=self.shift,
nb_melfilters=self.nb_melfilters,
nb_ceps=self.nb_ceps,
get_spec=self.get_spec,
get_mspec=self.get_mspec,
get_mfcc=self.get_mfcc,
get_qspec=self.get_qspec,
get_phase=self.get_phase,
get_pitch=self.get_pitch,
get_vad=self.get_vad,
get_energy=self.get_energy,
get_delta=self.get_delta,
pitch_threshold=self.pitch_threshold,
pitch_fmax=self.pitch_fmax,
vad_smooth=self.vad_smooth,
vad_minlen=self.vad_minlen,
cqt_bins=self.cqt_bins,
fmin=self.fmin,
fmax=self.fmax,
sr_new=self.sr_new,
preemphasis=self.preemphasis,
center=self.center)
if features is not None:
ret.append(
(name,
[features[i[0]] for i in self.__features_properties]))
else:
msg = 'Ignore segments: %s, error: NaN values' % name
warnings.warn(msg)
# return the results as a generator
return (i for i in ret)
except Exception as e:
msg = 'Ignore file: %s, error: %s' % (audio_path, str(e))
import traceback
traceback.print_exc()
raise e
def transform(self, X, start=0., end=-1, channel=0):
"""
fs : int
original sample frequency of data (in case reading pcm file
we don't know the original sample frequency)
"""
get_mspec = False
get_spec = False
get_mfcc = False
get_pitch = False
if 'mspec' in self.feature_type:
get_mspec = True
if 'spec' in self.feature_type:
get_spec = True
if 'mfcc' in self.feature_type:
get_mfcc = True
if 'pitch' in self.feature_type:
get_pitch = True
# ====== Read audio ====== #
if isinstance(X, str) and os.path.exists(X):
X, orig_fs = speech.read(X)
elif isinstance(X, np.ndarray):
orig_fs = None
elif isinstance(X, Data):
X = X[:]
orig_fs = None
else:
raise ValueError('Cannot process data type %s' % type(X))
# if specifed DEFAULT_FS
if orig_fs is None:
orig_fs = (SpeechTransform.DEFAULT_FS
if SpeechTransform.DEFAULT_FS is not None else self.fs)
# ====== check if downsample necessary ====== #
if self.fs < orig_fs: # downsample
from scikits.samplerate import resample
X = resample(X, self.fs / orig_fs, 'sinc_best')
elif self.fs > orig_fs:
raise ValueError('Cannot perform upsample from frequency: '
'{}Hz to {}Hz'.format(orig_fs, self.fs))
fs = orig_fs if self.fs is None else self.fs
# ====== preprocessing ====== #
N = len(X)
start = int(float(start) * fs)
end = int(N if end < 0 else end * fs)
X = X[start:end, channel] if X.ndim > 1 else X[start:end]
data = speech_features_extraction(X.ravel(),
fs=fs,
n_filters=self.n_filters,
n_ceps=self.n_ceps,
win=self.win,
shift=self.shift,
delta_order=self.delta_order,
energy=self.energy,
vad=self.vad,
dtype='float32',
pitch_threshold=self.pitch_threshold,
get_pitch=get_pitch,
get_spec=get_spec,
get_mspec=get_mspec,
get_mfcc=get_mfcc)
# ====== return results ====== #
if data is None:
return None
results = {}
if get_spec:
X, sum1, sum2 = data[0]
results['spec'] = X
if get_mspec:
X, sum1, sum2 = data[1]
results['mspec'] = X
if get_mfcc:
X, sum1, sum2 = data[2]
results['mfcc'] = X
if get_pitch:
X, sum1, sum2 = data[3]
results['pitch'] = X
results = [results[i] for i in self.feature_type]
if len(results) == 1: results = results[0]
if self.vad:
return (results +
[data[-1]] if isinstance(results,
(tuple,
list)) else [results, data[-1]])
return results
return results[0] if len(results) == 1 else results
files = [
'/Users/trungnt13/tmp/20051026_180611_340_fsp-b.sph',
'/Users/trungnt13/tmp/20051118_212058_553_fsp-b.sph',
'/Users/trungnt13/tmp/20110213_140144_892-b.sph',
'/Users/trungnt13/tmp/en_4077-a.sph',
'/Users/trungnt13/tmp/en_4660-a.sph',
'/Users/trungnt13/tmp/en_6402-b.sph',
'/Users/trungnt13/tmp/fla_0811-b.sph',
'/Users/trungnt13/tmp/fla_0946-b.sph',
'/Users/trungnt13/tmp/lre17_vlujcseb.sph'
]
for f in files:
print(f)
s, sr = speech.read(f, remove_dc_offset=True, remove_zeros=True)
segs, vad, voices, cut = test_func(s, sr, frame_length=0.02 * sr, maximum_duration=10, minimum_duration=None,
return_vad=True, return_cut=True, return_voices=True)
for s in segs:
print(s.shape, len(s) / sr)
plt.figure(figsize=(10, 6))
plt.subplot(4, 1, 1)
plt.plot(speech.resample(s, sr, 2000, best_algorithm=False))
plt.subplot(4, 1, 2)
plt.plot(vad)
plt.subplot(4, 1, 3)
plt.plot(voices)
plt.subplot(4, 1, 4)
plt.plot(cut)
plt.show()
import matplotlib
matplotlib.use('TkAgg')
from matplotlib import pyplot as plt
import seaborn
import numpy as np
import shutil
import os
from odin import fuel as F, utils
from odin.preprocessing import speech
from odin import visual
datapath = F.load_digit_wav()
print(datapath)
files = utils.get_all_files(datapath, lambda x: '.wav' in x)
y, sr = speech.read(files[0])
print('Raw signal:', y.shape, sr)
feat = speech.speech_features(y,
sr,
win=0.02,
shift=0.01,
nb_melfilters=40,
nb_ceps=13,
get_spec=True,
get_mspec=True,
get_mfcc=True,
get_qspec=True,
get_phase=True,
get_pitch=True,
get_vad=True,
return results[0] if len(results) == 1 else results
files = [
'/Users/trungnt13/tmp/20051026_180611_340_fsp-b.sph',
'/Users/trungnt13/tmp/20051118_212058_553_fsp-b.sph',
'/Users/trungnt13/tmp/20110213_140144_892-b.sph',
'/Users/trungnt13/tmp/en_4077-a.sph', '/Users/trungnt13/tmp/en_4660-a.sph',
'/Users/trungnt13/tmp/en_6402-b.sph',
'/Users/trungnt13/tmp/fla_0811-b.sph',
'/Users/trungnt13/tmp/fla_0946-b.sph',
'/Users/trungnt13/tmp/lre17_vlujcseb.sph'
]
for f in files:
print(f)
s, sr = speech.read(f, remove_dc_offset=True, remove_zeros=True)
segs, vad, voices, cut = test_func(s,
sr,
frame_length=0.02 * sr,
maximum_duration=10,
minimum_duration=None,
return_vad=True,
return_cut=True,
return_voices=True)
for s in segs:
print(s.shape, len(s) / sr)
plt.figure(figsize=(10, 6))
plt.subplot(4, 1, 1)
plt.plot(speech.resample(s, sr, 2000, best_algorithm=False))
plt.subplot(4, 1, 2)
plt.plot(vad)
def plot_audio(s,
sr=None,
win=0.02,
shift=0.01,
nb_melfilters=40,
nb_ceps=12,
get_qspec=False,
get_vad=True,
fmin=64,
fmax=None,
sr_new=None,
preemphasis=0.97,
pitch_threshold=0.8,
pitch_fmax=1200,
vad_smooth=3,
vad_minlen=0.1,
cqt_bins=96,
center=True,
title=""):
from matplotlib import pyplot as plt
from odin.preprocessing import speech
# ====== helper ====== #
def spectrogram(spec, vad, title):
plt.figure()
if spec.shape[0] / spec.shape[1] >= 8.:
nb_samples = len(spec)
n1, n2, n3 = nb_samples // 4, nb_samples // 2, 3 * nb_samples // 4
plt.subplot2grid((3, 4), (0, 0), rowspan=1, colspan=4)
plot_spectrogram(spec.T, vad=vad)
plt.subplot2grid((3, 4), (1, 0), rowspan=1, colspan=2)
plot_spectrogram(spec[:n1].T, vad=vad[:n1])
plt.subplot2grid((3, 4), (1, 2), rowspan=1, colspan=2)
plot_spectrogram(spec[n1:n2].T, vad=vad[n1:n2])
plt.subplot2grid((3, 4), (2, 0), rowspan=1, colspan=2)
plot_spectrogram(spec[n2:n3].T, vad=vad[n2:n3])
plt.subplot2grid((3, 4), (2, 2), rowspan=1, colspan=2)
plot_spectrogram(spec[n3:].T, vad=vad[n3:])
else:
plot_spectrogram(spec.T, vad=vad)
plt.suptitle(str(title))
plt.tight_layout()
# ====== load signal ====== #
if isinstance(s, string_types):
name = os.path.basename(s)
s, _ = speech.read(s)
if sr is None:
sr = _
else:
name = "-".join([str(s.shape), str(sr)])
title = str(title) + ":" + name
# ====== processing ====== #
get_vad = True if not get_vad else get_vad
y = speech.speech_features(s,
sr,
win=win,
shift=shift,
nb_melfilters=nb_melfilters,
nb_ceps=nb_ceps,
get_spec=True,
get_mspec=True,
get_mfcc=True,
get_qspec=get_qspec,
get_phase=False,
get_pitch=False,
get_vad=get_vad,
get_energy=True,
get_delta=False,
fmin=fmin,
fmax=fmax,
sr_new=sr_new,
preemphasis=preemphasis,
pitch_threshold=pitch_threshold,
pitch_fmax=pitch_fmax,
vad_smooth=vad_smooth,
vad_minlen=vad_minlen,
cqt_bins=cqt_bins,
center=center)
# ====== plot raw signals ====== #
if sr > 16000:
s = speech.resample(s, sr, 16000)
sr = 16000
plt.figure()
plt.subplot(2, 1, 1)
plt.plot(s)
plt.subplot(2, 1, 2)
plt.plot(y['energy'].ravel())
plt.tight_layout()
plt.suptitle(title)
# ====== plot spectrogram ====== #
spectrogram(y['spec'], y['vad'], title='STFT power spectrum')
if get_qspec:
spectrogram(y['qspec'], y['vad'], title='CQT power spectrum')
return y