def extract_features(self):
# Generates a Short-time Fourier transform (STFT) to use in the chroma_stft
stft = np.abs(librosa.stft(self.waveform))
# Generate Mel-frequency cepstral coefficients (MFCCs) from a time series
mfccs = mfcc(y=self.waveform, sr=self.sample_rate,
n_mfcc=40).mean(axis=1)
# Computes a chromagram from a waveform or power spectrogram.
chroma = chroma_stft(S=stft, sr=self.sample_rate).mean(axis=1)
# Computes a mel-scaled spectrogram.
mel = melspectrogram(self.waveform, sr=self.sample_rate).mean(axis=1)
# Computes spectral contrast
contrast = spectral_contrast(S=stft, sr=self.sample_rate).mean(axis=1)
# Computes the tonal centroid features (tonnetz)
harmonic = librosa.effects.harmonic(self.waveform)
tonn = tonnetz(y=harmonic, sr=self.sample_rate).mean(axis=1)
return np.concatenate([mfccs, chroma, mel, contrast, tonn], axis=0)
def _calc_feat(self, window, feat_name):
feat = None
# calculate feature
if feat_name == 'mfcc':
feat = FT.mfcc(y=window, sr=self.sr, n_mfcc=_N_MFCC)
elif feat_name == 'chroma_stft':
feat = FT.chroma_stft(y=window, sr=self.sr)
elif feat_name == 'melspectrogram':
feat = FT.melspectrogram(y=window,
sr=self.sr,
n_mels=128,
n_fft=1024,
hop_length=512)
feat = L.power_to_db(feat)
elif feat_name == 'spectral_centroid':
feat = FT.spectral_centroid(y=window, sr=self.sr)
elif feat_name == 'spectral_rolloff':
feat = FT.spectral_rolloff(y=window, sr=self.sr)
elif feat_name == 'tonnetz':
feat = FT.tonnetz(y=window, sr=self.sr)
elif feat_name == 'zero_crossing_rate':
feat = FT.zero_crossing_rate(y=window)
else:
assert False, 'Invalid feature'
# pool feature from multiple frames
if self.feature_pool == 'sum':
feat = feat.sum(axis=1)
elif self.feature_pool == 'max':
feat = feat.max(axis=1)
elif self.feature_pool == 'mean':
feat = feat.mean(axis=1)
elif self.feature_pool == 'flatten':
feat = feat.flatten()
elif self.feature_pool == 'none':
pass
else:
assert False, 'Invalid feature pooling scheme'
# normalize features
if self.l2_norm and feat.shape[0] > 1:
feat /= np.linalg.norm(feat)
return feat
def get_tonal_centroid_transform(y, sr, tonal_model, doce_bins_tuned_chroma):
"""
returns centroids from tonal model
Parameters
----------
hpss : bool
true or false depends on hpss block
name_file: str
name of the file that is being computed
y : number > 0 [scalar]
audio
sr: number > 0 [scalar]
target sampling rate
chroma: str
chroma-samplerate-framesize-overlap
tonal_model: str optional
Tonal model block type. "TIV2" for Tonal Interval space focus on audio. "TIV2" for audio. "TIV2_Symb" for symbolic data.
"tonnetz" for harte centroids aproach. Default TIV2\
doce_bins_tuned_chroma: list
list of chroma vectors
Returns
-------
list of tonal centroids vectors
"""
centroid_vector = None
if tonal_model == 'tonnetz':
centroid_vector = tonnetz(y=y, sr=sr, chroma=doce_bins_tuned_chroma)
elif tonal_model == 'TIV2':
centroid_vector = tonal_interval_space(doce_bins_tuned_chroma)
elif tonal_model == 'TIV2_symb':
centroid_vector = tonal_interval_space(doce_bins_tuned_chroma,
symbolic=True)
return centroid_vector
y, sr = load(my_file)
D = np.abs(stft(y))
specshow(amplitude_to_db(D, ref=np.max), y_axis='log', x_axis='time')
plt.title("Power Spectrogram")
plt.colorbar(format='%+2.0f dB')
plt.tight_layout()
chroma_cq = chroma_cqt(y=y, sr=sr)
specshow(chroma_cq, y_axis='chroma', x_axis='time')
plt.title("Chromagram Constant Q Transform")
plt.colorbar()
plt.tight_layout()
tonnetz = tonnetz(y=y, sr=sr)
specshow(tonnetz, y_axis='tonnetz')
plt.title("Tonnetz Example")
plt.colorbar()
plt.tight_layout()
ms = melspectrogram(y=y, sr=sr)
specshow(power_to_db(ms, ref=np.max), y_axis='mel', fmax=8000, x_axis='time')
plt.title("Mel Spectrogram Example")
plt.colorbar(format="%+2.0f dB")
plt.tight_layout()
waveplot(y=y, sr=sr)
plt.title("Audio Waveplot")
plt.tight_layout()
def tonnetz(args):
sig = get_sig(args)
fs = args['fs']
return rosaft.tonnetz(y=sig, sr=fs)
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)
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))
row = np.append(row, thing1[-1])
#print(len(row))
train_data = np.append(train_data, row)
counter += 1
columns = ["feat_" + str(i) for i in range(299)]
columns.append("class")
def tonnetz(self):
sig = get_sig(self.args)
fs = self.args['fs']
return rosaft.tonnetz(y=sig, sr=fs)
chroma-samplerate-framesize-overlap tonal_model: str optional Tonal model block type. "TIV2" for Tonal Interval space focus on audio. "TIV2" for audio. "TIV2_Symb" for symbolic data. "tonnetz" for harte centroids aproach. Default TIV2\ doce_bins_tuned_chroma: list list of chroma vectors Returns ------- list of tonal centroids vectors """ centroid_vector = None if tonal_model == 'tonnetz': centroid_vector = tonnetz(y=y, sr=sr, chroma=doce_bins_tuned_chroma) elif tonal_model == 'TIV2': centroid_vector = tonal_interval_space(doce_bins_tuned_chroma) elif tonal_model == 'TIV2_symb': centroid_vector = tonal_interval_space(doce_bins_tuned_chroma, symbolic=True) return centroid_vector def tonal_centroid_transform(hpss, chroma, name_file, y, sr, tonal_model, doce_bins_tuned_chroma): """ wrapper of tonal centroid transform for save all results for future same calculations Parameters ---------- hpss : bool true or false depends on hpss block