def test_spectral_flatness(self):
correct = rosaft.spectral_flatness(y=self.sig,
S=None,
n_fft=nfft,
hop_length=stepsize)
actual = spectral_flatness(self.args)
self.assertTrue(np.abs(correct - actual).max() < tol)
def spectral_flatness(args):
psd = get_psd(args)
nfft, noverlap = unroll_args(args, ['nfft', 'noverlap'])
hopsize = nfft - noverlap
return rosaft.spectral_flatness(y=None,
S=psd,
n_fft=nfft,
hop_length=hopsize)
def extract_feature(self, audio_data):
"""
extract features from audio data
:param audio_data:
:return:
"""
zcr = lrf.zero_crossing_rate(audio_data,
frame_length=self.FRAME,
hop_length=self.FRAME / 2)
feature_zcr = np.mean(zcr)
ste = audio_utils.AudioUtils.ste(audio_data, 'hamming',
int(20 * 0.001 * self.RATE))
feature_ste = np.mean(ste)
ste_acc = np.diff(ste)
feature_steacc = np.mean(ste_acc[ste_acc > 0])
stzcr = audio_utils.AudioUtils.stzcr(audio_data, 'hamming',
int(20 * 0.001 * self.RATE))
feature_stezcr = np.mean(stzcr)
mfcc = lrf.mfcc(y=audio_data, sr=self.RATE, n_mfcc=13)
feature_mfcc = np.mean(mfcc, axis=1)
spectral_centroid = lrf.spectral_centroid(y=audio_data,
sr=self.RATE,
hop_length=self.FRAME / 2)
feature_spectral_centroid = np.mean(spectral_centroid)
spectral_bandwidth = lrf.spectral_bandwidth(y=audio_data,
sr=self.RATE,
hop_length=self.FRAME / 2)
feature_spectral_bandwidth = np.mean(spectral_bandwidth)
spectral_rolloff = lrf.spectral_rolloff(y=audio_data,
sr=self.RATE,
hop_length=self.FRAME / 2,
roll_percent=0.90)
feature_spectral_rolloff = np.mean(spectral_rolloff)
spectral_flatness = lrf.spectral_flatness(y=audio_data,
hop_length=self.FRAME / 2)
feature_spectral_flatness = np.mean(spectral_flatness)
features = np.append([
feature_zcr, feature_ste, feature_steacc, feature_stezcr,
feature_spectral_centroid, feature_spectral_bandwidth,
feature_spectral_rolloff, feature_spectral_flatness
], feature_mfcc)
return features, self.label
def get_mir(audio_path):
hop_length = 200
# Spectral Flux/Flatness, MFCCs, SDCs
spectrogram = madmom.audio.spectrogram.Spectrogram(audio_path,
frame_size=2048,
hop_size=hop_length,
fft_size=4096)
# only take 30s snippets to align data
audio = madmom.audio.signal.Signal(audio_path,
dtype=float,
start=0,
stop=30)
all_features = []
#print(spectrogram.shape)
#print(audio.shape)
#print('signal sampling rate: {}'.format(audio.sample_rate))
# madmom features
all_features.extend([
spectral_flux(spectrogram),
superflux(spectrogram),
complex_flux(spectrogram)
]) #, MFCC(spectrogram)])
# mfcc still wrong shape as it is a 2 array
# librosa features
libr_features = [
spectral_centroid(audio, hop_length=hop_length),
spectral_bandwidth(audio, hop_length=hop_length),
spectral_flatness(audio, hop_length=hop_length),
spectral_rolloff(audio, hop_length=hop_length),
rmse(audio, hop_length=hop_length),
zero_crossing_rate(audio, hop_length=hop_length)
] #, mfcc(audio)])
for libr in libr_features:
all_features.append(np.squeeze(libr, axis=0))
# for feature in all_features:
# print(feature.shape)
X = np.stack(all_features, axis=1)[na, :, :]
return X
def feature_extraction_all(signal, sr, n_mfcc, buffer_len,
normalization_values):
"""
Feature extraction interface
:param signal: Signal
:param sr: Signal
:param n_mfcc: Signal
:param buffer_len: Signal
:param normalization_values: normalization values of the dataset
:output features: Array of features
Features are extracted from the incoming audio signal when an onset is detected.
"""
features = []
signal = np.array(signal)
if signal.size != 0:
S, phase = librosa.magphase(
librosa.stft(y=signal,
n_fft=buffer_len,
hop_length=int(buffer_len / 4)))
# Mel Frequency cepstral coefficients
mfcc = feature.mfcc(y=signal,
sr=sr,
n_mfcc=n_mfcc,
n_fft=int(512 * 2),
hop_length=int(128 * 2))
mfcc_mean = np.mean(mfcc, axis=1)
mfcc_std = np.std(mfcc, axis=1)
# RMS
rms = feature.rms(S=S,
frame_length=buffer_len,
hop_length=int(buffer_len / 4))
rms_mean = np.mean(rms, axis=1)
rms_std = np.std(rms, axis=1)
# Spectral Centroid
spectral_centroid = feature.spectral_centroid(S=S, sr=sr)
spectral_centroid_mean = np.mean(spectral_centroid, axis=1)
spectral_centroid_std = np.std(spectral_centroid, axis=1)
# Rolloff
spectral_rolloff = feature.spectral_rolloff(S=S, sr=sr)
spectral_rolloff_mean = np.mean(spectral_rolloff, axis=1)
spectral_rolloff_std = np.std(spectral_rolloff, axis=1)
# Bandwidth
spectral_bandwidth = feature.spectral_bandwidth(S=S, sr=sr)
spectral_bandwidth_mean = np.mean(spectral_bandwidth, axis=1)
spectral_bandwidth_std = np.std(spectral_bandwidth, axis=1)
# Contrast
spectral_contrast = feature.spectral_contrast(S=S, sr=sr)
spectral_contrast_mean = np.mean(spectral_contrast, axis=1)
spectral_contrast_std = np.std(spectral_contrast, axis=1)
# Flatness
spectral_flatness = feature.spectral_flatness(S=S)
spectral_flatness_mean = np.mean(spectral_flatness, axis=1)
spectral_flatness_std = np.std(spectral_flatness, axis=1)
if len(normalization_values) > 1:
# Duration
features.append(
normalize(len(signal), normalization_values['duration']))
features.extend(
normalize(
mfcc_mean, normalization_values[[
'mfcc_mean_1', 'mfcc_mean_2', 'mfcc_mean_3',
'mfcc_mean_4', 'mfcc_mean_5', 'mfcc_mean_6',
'mfcc_mean_7', 'mfcc_mean_8', 'mfcc_mean_9',
'mfcc_mean_10'
]]))
features.extend(
normalize(
mfcc_std, normalization_values[[
'mfcc_std_1', 'mfcc_std_2', 'mfcc_std_3', 'mfcc_std_4',
'mfcc_std_5', 'mfcc_std_6', 'mfcc_std_7', 'mfcc_std_8',
'mfcc_std_9', 'mfcc_std_10'
]]))
features.extend(
normalize(rms_mean, normalization_values['rms_mean']))
features.extend(normalize(rms_std,
normalization_values['rms_std']))
features.extend(
normalize(spectral_centroid_mean,
normalization_values['spectral_centroid_mean']))
features.extend(
normalize(spectral_centroid_std,
normalization_values['spectral_centroid_std']))
features.extend(
normalize(spectral_rolloff_mean,
normalization_values['spectral_rolloff_mean']))
features.extend(
normalize(spectral_rolloff_std,
normalization_values['spectral_rolloff_std']))
features.extend(
normalize(spectral_bandwidth_mean,
normalization_values['spectral_bandwidth_mean']))
features.extend(
normalize(spectral_bandwidth_std,
normalization_values['spectral_bandwidth_std']))
features.extend(
normalize(
spectral_contrast_mean, normalization_values[[
'spectral_contrast_mean_1', 'spectral_contrast_mean_2',
'spectral_contrast_mean_3', 'spectral_contrast_mean_4',
'spectral_contrast_mean_5', 'spectral_contrast_mean_6',
'spectral_contrast_mean_7'
]]))
features.extend(
normalize(
spectral_contrast_std, normalization_values[[
'spectral_contrast_std_1', 'spectral_contrast_std_2',
'spectral_contrast_std_3', 'spectral_contrast_std_4',
'spectral_contrast_std_5', 'spectral_contrast_std_6',
'spectral_contrast_std_7'
]]))
features.extend(
normalize(spectral_flatness_mean,
normalization_values['spectral_flatness_mean']))
features.extend(
normalize(spectral_flatness_std,
normalization_values['spectral_flatness_std']))
else:
features.append(len(signal))
features.extend(mfcc_mean)
features.extend(mfcc_std)
features.extend(rms_mean)
features.extend(rms_std)
features.extend(spectral_centroid_mean)
features.extend(spectral_centroid_std)
features.extend(spectral_rolloff_mean)
features.extend(spectral_rolloff_std)
features.extend(spectral_bandwidth_mean)
features.extend(spectral_bandwidth_std)
features.extend(spectral_contrast_mean)
features.extend(spectral_contrast_std)
features.extend(spectral_flatness_mean)
features.extend(spectral_flatness_std)
features = np.array(features)
return features
def extract_features(soundwave,sampling_rate,sound_name="test",feature_list=[]):
"""
extracts features with help of librosa
:param soundwave: extracted soundwave from file
:param sampling_rate: sampling rate
:param feature_list: list of features to compute
:param sound_name: type of sound, i.e. dog
:return: np.array of all features for the soundwave
"""
print("Computing features for ",sound_name)
if len(feature_list)==0:
feature_list=["chroma_stft","chroma_cqt","chroma_cens","melspectrogram",
"mfcc","rmse","spectral_centroid","spectral_bandwidth",
"spectral_contrast","spectral_flatness","spectral_rolloff",
"poly_features","tonnetz","zero_crossing_rate"]
features=[]
#feature_len
#"chroma_stft":12
if "chroma_stft" in feature_list:
features.append(feat.chroma_stft(soundwave, sampling_rate))
#"chroma_cqt":12
if "chroma_cqt" in feature_list:
features.append(feat.chroma_cqt(soundwave, sampling_rate))
#"chroma_cens":12
if "chroma_cens" in feature_list:
features.append(feat.chroma_cens(soundwave, sampling_rate))
#"malspectrogram":128
if "melspectrogram" in feature_list:
features.append(feat.melspectrogram(soundwave, sampling_rate))
#"mfcc":20
if "mfcc" in feature_list:
features.append(feat.mfcc(soundwave, sampling_rate))
#"rmse":1
if "rmse" in feature_list:
features.append(feat.rmse(soundwave))
#"spectral_centroid":1
if "spectral_centroid" in feature_list:
features.append(feat.spectral_centroid(soundwave, sampling_rate))
#"spectral_bandwidth":1
if "spectral_bandwidth" in feature_list:
features.append(feat.spectral_bandwidth(soundwave, sampling_rate))
#"spectral_contrast":7
if "spectral_contrast" in feature_list:
features.append(feat.spectral_contrast(soundwave, sampling_rate))
#"spectral_flatness":1
if "spectral_flatness" in feature_list:
features.append(feat.spectral_flatness(soundwave))
#"spectral_rolloff":1
if "spectral_rolloff" in feature_list:
features.append(feat.spectral_rolloff(soundwave, sampling_rate))
#"poly_features":2
if "poly_features" in feature_list:
features.append(feat.poly_features(soundwave, sampling_rate))
#"tonnetz":6
if "tonnetz" in feature_list:
features.append(feat.tonnetz(soundwave, sampling_rate))
#"zero_crossing_rate":1
if "zero_crossing_rate" in feature_list:
features.append(feat.zero_crossing_rate(soundwave))
return np.concatenate(features)
#print(chunk)
chunk1 = np.array(chunk)
for thing in chunk1:
print(counter)
thing1 = np.array(thing)
#print(thing1)
row = np.array([])
cstft = np.mean(lf.chroma_stft(thing1[:-1]).T, axis=0)
row = np.concatenate((row, cstft))
cqt = np.mean(lf.chroma_cqt(thing1[:-1]).T, axis=0)
row = np.concatenate((row, cqt))
sens = np.mean(lf.chroma_cens(thing1[:-1]).T, axis=0)
row = np.concatenate((row, sens))
spcent = np.mean(lf.spectral_centroid(thing1[:-1]).T, axis=0)
row = np.concatenate((row, spcent))
flatness = np.mean(lf.spectral_flatness(thing1[:-1]).T, axis=0)
row = np.concatenate((row, flatness))
rolloff = np.mean(lf.spectral_rolloff(thing1[:-1]).T, axis=0)
row = np.concatenate((row, rolloff))
mspec = np.mean(lf.melspectrogram(thing1[:-1]).T, axis=0)
row = np.concatenate((row, mspec))
mfcc = np.mean(lf.mfcc(thing1[:-1], n_mfcc=30).T, axis=0)
row = np.concatenate((row, mfcc))
tonnetz = np.mean(lf.tonnetz(thing1[:-1]).T, axis=0)
row = np.concatenate((row, tonnetz))
rmse = np.mean(lf.rmse(thing1[:-1]).T, axis=0)
row = np.concatenate((row, rmse))
contrast = np.mean(lf.spectral_contrast(thing1[:-1]).T, axis=0)
row = np.concatenate((row, contrast))
tempo = np.mean(lf.tempogram(thing[:-1], win_length=88).T, axis=0)
row = np.concatenate((row, tempo))
def compute_average_spectral_flatness(y):
return np.mean(spectral_flatness(y=y))
def flatness(wave_form, sample_rate, hop_length):
return feature.spectral_flatness(y=wave_form, hop_length=hop_length).T