def melspectrogram(self):
print(self.directory)
rate = self.rate
seg_length = 2 * (dim - 1)
fft = int((2.99936669e-02 * self.data.shape[0]) + 1.42217233e+02)
hop = int((0.01516035 * self.data.shape[0]) - 2.14982993)
S = melspectrogram(self.data,
self.rate,
n_mels=mels,
hop_length=hop,
center=False,
n_fft=fft,
norm=1,
fmax=fmax_,
power=poww)
while S.shape[1] != dim:
if S.shape[1] > dim:
fft += 2
elif S.shape[1] < dim:
hop -= 1
S = melspectrogram(self.data,
self.rate,
n_mels=mels,
hop_length=hop,
n_fft=fft,
center=False,
norm=1,
fmax=fmax_,
power=poww)
S = 255 * (S - S.min()) / (S.max() - S.min())
return (S)
def transform_audio(self, y):
'''Compute the PCEN of the (log-) Mel Spectrogram
Parameters
----------
y : np.ndarray
The audio buffer
Returns
-------
data : dict
data['mag'] : np.ndarray, shape = (n_frames, n_bins)
The PCEN magnitude
'''
#extract proper shape
S_test = melspectrogram(y=y,
sr=self.sr,
hop_length=self.hop_length,
n_fft=self.n_fft,
n_mels=self.n_mels)
P_test = pcen(S_test,
sr=self.sr,
hop_length=self.hop_length,
time_constant=1)
n_frames = P_test.shape[1]
#double audio and reverse pad to prevent zero initial-energy assumption
y = np.concatenate((y[::-1], y))
S = melspectrogram(y=y,
sr=self.sr,
hop_length=self.hop_length,
n_fft=self.n_fft,
n_mels=self.n_mels)
if self.log:
S = amplitude_to_db(S, ref=np.max)
t_base = (self.hop_length) / (self.sr) #tau, or hop length in time
t_constants = t_base * np.array(
[2**i for i in range(self.n_t_constants)])
pcen_layers = []
for T in t_constants:
P = pcen(S,
sr=self.sr,
hop_length=self.hop_length,
time_constant=T)
#source of off-by-one error:
P = P[:, P.shape[1] - n_frames + 1:] #remove padded section
P = to_dtype(P, self.dtype)
pcen_layers.append(P)
pcen_layers = to_dtype(np.asarray(pcen_layers), self.dtype)
return {
'mag': self._index(pcen_layers)
} #copied from mel spectrogram pump feature extractor
class ExtractMonoAudioFiles(FeaturesExtractor):
#static vars to be manually set
##the natural byterate of the examples here but we can consider modifying it
sr = 44100
nblabels = 88
batchsize = 1000
#featurefunc = lambda y, sr: lrft.mfcc(y, sr, n_mfcc=1).T
#featurefunc = lambda *x: x
featurefunc = lambda y, sr: lrft.melspectrogram(y, sr).T
inpath = '../simple-wdir'
#for feeder
featuremutation = lambda y, sr: lrft.melspectrogram(y, sr).T
@staticmethod
def labelmutation(pitch, nbsamples):
labelvect = np.zeros(shape=(nbsamples, ExtractMonoAudioFiles.nblabels))
labelvect[:, int(pitch)] = np.ones(nbsamples)
return labelvect
def __init__(self, inpath=None):
if inpath is None:
self.inpath = ExtractMonoAudioFiles.inpath
else:
self.inpath = inpath
# if featurefunc is None:
# self.featurefunc = ExtractMonoAudioFiles.featurefunc
# else:
# self.featurefunc = featurefunc
super().__init__(self.inpath, ExtractMonoAudioFiles.computefeatureddata, ExtractMonoAudioFiles.extractsamplesmetadata)
@staticmethod
def extractsamplesmetadata(path, filelist):
samplesmetadata = []
for f in filelist:
with open(join(path, f), 'r') as fs:
#get the the onset and offset plus midi pitch
onset, offset, midipitch = tuple(map(float, fs.readlines()[1][:-1].split('\t')))
#first is the midi pitch rescaled into [0,87] range
#second is a tuple of wav path, onset and duration
samplesmetadata.append(((f[:-3] + 'wav', onset, offset-onset), int(midipitch - 21)))
return samplesmetadata
@staticmethod
def computefeatureddata(path, samplemetadata):
meta, pitch = samplemetadata
audiodat = lrco.load(join(path, meta[0]), sr= ExtractMonoAudioFiles.sr,
offset=meta[1], duration=meta[2])
audiodat = ExtractMonoAudioFiles.featurefunc(*audiodat)
pitchvect = np.array([pitch] * audiodat.shape[0])
#return (audiodat, pitch)
return (audiodat, pitchvect)
def logmel(self):
from librosa.feature import melspectrogram
from librosa.core import load
logmel_params = self.config['logmel_params']
sr = logmel_params['sr']
n_fft = logmel_params['n_fft']
hop_length = logmel_params['hop_length']
n_mels = logmel_params['n_mels']
feature_path = os.path.join(
self.dataset['feature_path'],
'logmel_{}_{}_{}_{}'.format(sr, n_fft, hop_length, n_mels))
if not os.path.exists(feature_path):
os.mkdir(feature_path)
x_train = []
y_train = []
f_train = []
for i, row in self.dataset.train_data.iterrows():
print('[Train] {}) Getting logmels from {}...'.format(
i, row['cur_name']),
end='')
wav_name = os.path.join(self.dataset['data_path'], row['cur_name'])
wav_data, sr = load(wav_name, sr=sr)
x_train.append(
melspectrogram(wav_data,
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
n_mels=n_mels))
y_train.append(self._build_multilabel(row))
f_train.append(row['cur_name'])
print('done.')
x_test = []
y_test = []
f_test = []
for i, row in self.dataset.test_data.iterrows():
print('[Test] {}) Getting mels from {}...'.format(
i, row['cur_name']),
end='')
wav_name = os.path.join(self.dataset['data_path'], row['cur_name'])
wav_data, sr = load(wav_name, sr=sr)
x_test.append(
melspectrogram(wav_data,
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
n_mels=n_mels))
y_test.append(self._build_multilabel(row))
f_test.append(row['cur_name'])
print('done')
self._save_pickles(feature_path, x_train, y_train, f_train, x_test,
y_test, f_test)
def get_mult_feature(wave_name, window):
(rate, sig) = wav.read(wave_name) # librosa.load()
#
#power = librosa.amplitude_to_db(librosa.stft(sig,n_fft = window,hop_length=window/2), ref=numpy.max)
power = numpy.abs(librosa.stft(sig, n_fft=window, hop_length=window / 2))
_min = numpy.min(power)
_max = numpy.max(power)
power = (power - _min) / (_max - _min)
print power.shape
mels = feature.melspectrogram(sig,
rate,
n_fft=window,
hop_length=window / 2,
n_mels=257)
_min = numpy.min(mels)
_max = numpy.max(mels)
mels = (mels - _min) / (_max - _min)
print mels.shape
if power.shape[1] < 250:
zero = numpy.zeros((power.shape[0], 250 - power.shape[1]))
power = numpy.hstack((power, zero))
mels = numpy.hstack((mels, zero))
else:
power = power[:, 0:250]
mels = mels[:, 0:250]
power = power.T
mels = mels.T
feat = numpy.asarray([power, mels])
return feat
def wav_to_mel_util(x):
mel_spec = melspectrogram(x,
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
window=window)
return mel_spec
def process(src_fname_path, dest_fname_path):
if not src_fname_path.endswith('.mp3'):
return
try:
# load and process audio
audio, sr = librosa.load(src_fname_path)
# audio = lowpass(audio, cutoff=3000, sample_freq=sr)
# spec = librosa.stft(np.asfortranarray(audio))
spec = melspectrogram(audio, sr)
spec_db = librosa.amplitude_to_db(np.abs(spec))
# generate plot
scale = 1
fig = plt.figure(figsize=(1.28 * scale, 0.64 * scale)) #128x64
plt.box(False)
plt.subplots_adjust(left=0, right=1, bottom=0, wspace=0, hspace=0)
librosa.display.specshow(spec_db,
sr=sr,
cmap='gray_r',
x_axis='time',
y_axis='log')
fig.savefig(dest_fname_path, bbox_inches=None, pad_inches=0)
plt.close()
print('{0} -> {1}'.format(src_fname_path, dest_fname_path))
except Exception as e:
print('processing {0}: {1}'.format(src_fname_path, e))
def model_predict():
"""
Prediction of the classes 'open' and 'close'
"""
# Create a numpy array of audio data
signal = np.frombuffer(stream.read(config.new_len,
exception_on_overflow=False),
dtype=np.float32)
# signal_merged = np.append(config.signal_old, signal)
# config.signal_old = signal_merged[signal_len:]
signal_merged = signal
# Using random signal from folder 'clean' (turned off)
# signal_merged = evaluation_using_random()
mel = melspectrogram(y=signal_merged,
sr=config.sample_rate,
n_mels=config.n_mels,
n_fft=config.n_fft,
hop_length=config.hop_length,
window=config.window)
X = librosa.amplitude_to_db(mel)
X = X.reshape(1, X.shape[0], X.shape[1], 1)
# Prediction
prediction = model.predict([X])
# start_time = time.time()
# Show the prediction, signal and mel (turned off)
signal_visualisation(prediction, mel, signal_merged)
# print(str((time.time() - start_time)*1000)+'ms')
return prediction[0][0], prediction[0][1]
def process_individual_file(input_file_uri, index):
"""
It processes each input file and saves those in the required
format, which is going to be used while training.
Inputs:
input_file_name : name of the data file to be used during training.
Returns:
{output_file_uri, num_of_mel_frames, length_of_time_scaled_audio}
"""
data = [librosa_import(uri)[0] for uri in files_uri][0]
mfcc = melspectrogram(data,
n_fft=params.nFft,
hop_length=params.hop_size,
n_mels=params.num_mels).T
mfcc_shape = mfcc.shape
num_elem_in_mfcc = mfcc_shape[0] * mfcc_shape[1]
pad_val = params.scale_factor * num_elem_in_mfcc - data.shape[0]
data = np.pad(data, [0, pad_val], mode="constant")
assert data.shape[0] % num_elem_in_mfcc == 0
output_file_uri = os.path.join(training_data_folder,
"sample_{}".format(index))
postprocess_data(data, mfcc, output_file_uri)
return {output_file_uri, mfcc_shape[0], len(data)}
def compute_spectrogram(self):
# Compute the spectrogram and convert to dB
if not hasattr(self, "spectrogram"):
self.spectrogram = melspectrogram(self.waveform,
self.sampling_rate)
self.spectrogram_db = power_to_db(self.spectrogram, ref=np.max)
print("... Computed spectrogram")
def process_signal(self, signal):
ft = np.abs(stft(signal, n_fft=self.window_size, hop_length=self.window_stride, window='hann'))
mel = melspectrogram(sr=self.sample_rate,S=ft)
mfccs = mfcc( sr=self.sample_rate, n_mfcc=self.num_mfccs,S=mel)
deltas= delta(mfccs)
delta_deltas= delta(mfccs,order=2)
return mfccs, deltas, delta_deltas
def __data_generation(self, list_IDs_temp):
#'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)
# Initialization
X = np.empty((self.batch_size, *self.dim, self.n_channels))
Y = np.empty((self.batch_size, self.n_classes), dtype=np.bool)
# Generate data
for i, row in list_IDs_temp.iterrows():
if row.path not in self.audio.keys():
#print('{} - loading {}'.format(i, row.path))
#sys.stdout.flush()
aud, fs = load(row.path)
coefs = melspectrogram(aud,
sr=fs,
n_fft=2**12,
hop_length=2**11,
n_mels=64,
fmax=10000)
self.audio[row.path] = coefs
#print('{} - loaded!'.format(i))
#sys.stdout.flush()
# we've loaded one more track, add it to the counter
self.pbar.update(1)
start_ind = np.random.randint(low=0,
high=self.audio[row.path].shape[1] -
self.window)
clip = self.audio[row.path][:, start_ind:start_ind + self.window]
#
# start_ind = np.random.randint(low=0,high=coefs.shape[1]-self.window)
# clip = coefs[:,start_ind:start_ind+self.window]
X[i, :, :, 0] = clip
Y[i, :] = row.iloc[2:-1].values.astype(np.int64)
return X, Y
def wave2spec(
wave, fs, frame_period, window,
nperseg=None, nmels=80, preemphasis_coef=None,
f_min=0, f_max=None, dtype='float32',
return_t=False):
stft_kwargs = make_stft_args(
frame_period, fs, nperseg=nperseg, window=window)
htk, norm = True, "slaney"
if preemphasis_coef is not None:
spec_wave = preemphasis(wave, preemphasis_coef)
else:
spec_wave = wave
_, t, Zxx = signal.stft(
spec_wave, **stft_kwargs)
pspec = np.abs(Zxx)
mspec = melspectrogram(
sr=fs, S=pspec,
n_mels=nmels, fmin=f_min, fmax=f_max,
power=1.0, htk=htk, norm=norm)
pspec = pspec.T.astype(dtype)
mspec = mspec.T.astype(dtype)
upsample = fs // (1000 // frame_period)
length = (len(wave) // upsample) * upsample
wave = wave[:length]
mspec = mspec[:length // upsample]
spec = pspec[:length // upsample]
if return_t:
return wave, spec, mspec, upsample, t
else:
return wave, spec, mspec, upsample
def spectogram(audio):
spec = melspectrogram(audio, sr=audio_params['sr'], center=True,
n_mels=audio_params['n_mel'], n_fft=audio_params['n_fft'],
win_length=audio_params['win_length'], hop_length=audio_params['hop_length'],
fmin=audio_params['f_min'],
fmax=audio_params['f_max'])
spec = np.log(spec)
def extract_features(signal, normalize=False, wavelet=0):
# handle less than 3 [seg]
L = sr*3 # Total length for samples ~3[seg]
signal_length = signal.shape[0]
if signal_length < L:
#pad by repeating signal
signal = np.pad(signal, (0, L-signal_length), mode='wrap')
elif signal_length > L:
signal = signal[:L]
# Calculate melspectrogram
melspec = melspectrogram(signal, sr=sr, center=False, #fmax = sr/2
hop_length=window_size, win_length=window_size,
n_mels=128) # shape:[bands, frames]
# Transform to log scale and transpose
melspec = power_to_db(melspec, ref=np.amax(melspec)).T # shape:[frames, bands]
if normalize:
melspec = (melspec - np.mean(melspec))/np.std(melspec)
# 2D Discrete Wavelet Transform
if wavelet != 0:
LL, (LH, HL, HH) = dwt2(melspec, wavelet)
melspec = np.stack([LL,LH,HL,HH],axis=-1) # shape: [frames, bands, 4]
else:
melspec = melspec[..., np.newaxis]
# Reshape
features = melspec[np.newaxis, ...] # shape : [1, frames, bands, channels]
return features
def create_mels_deltas(waveform, sample_rate):
one_mel = melspectrogram(waveform.squeeze(0).numpy(),
sr=sample_rate,
n_fft=2048,
hop_length=1024,
n_mels=128,
fmin=0.0,
fmax=sample_rate / 2,
htk=True,
norm=None)
one_mel = np.log(one_mel + 1e-8)
one_mel = (one_mel - np.min(one_mel)) / (np.max(one_mel) - np.min(one_mel))
one_mel_delta = delta(one_mel)
one_mel_delta = (one_mel_delta - np.min(one_mel_delta)) / (
np.max(one_mel_delta) - np.min(one_mel_delta))
one_mel_delta_delta = delta(one_mel, order=2)
one_mel_delta_delta = (one_mel_delta_delta - np.min(one_mel_delta_delta)
) / (np.max(one_mel_delta_delta) -
np.min(one_mel_delta_delta))
mel_3d = torch.cat([
torch.tensor(one_mel).unsqueeze(0),
torch.tensor(one_mel_delta).unsqueeze(0),
torch.tensor(one_mel_delta_delta).unsqueeze(0)
],
dim=0)
return mel_3d
def prepare_data_streaminput(data, config):
X = np.empty(shape=(1, config.dim[0], config.dim[1], 1))
input_length = config.audio_length
# Remove silence or noise or things like that
# Random offset / Padding
if len(data) > input_length:
max_offset = len(data) - input_length
offset = np.random.randint(max_offset)
data = data[offset:(input_length + offset)]
else:
if input_length > len(data):
max_offset = input_length - len(data)
offset = np.random.randint(max_offset)
else:
offset = 0
data = np.pad(data, (offset, input_length - len(data) - offset),
"constant")
#data = librosa.feature.mfcc(data, sr=config.sampling_rate, n_mfcc=config.n_mfcc)
S = feature.melspectrogram(data,
sr=config.sampling_rate,
n_fft=2048,
hop_length=512,
n_mels=config.n_mfcc)
S_DB = power_to_db(S, ref=np.max)
data = np.expand_dims(S_DB, axis=-1)
X[0, ] = data
return X
def extract_audio_features(root_dir, row):
raw_data_dir = join(root_dir, RAW_DATA_DIR)
row_dict = row.to_dict()
waveform, _ = load(raw_data_dir + row_dict['filename'], sr=FEATURE_ARGS['sr'])
row_dict['melspec'] = _clean_features(melspectrogram(waveform, n_mels=EXTRACTOR_ARGS['n_mels'], **FEATURE_ARGS))
row_dict['mfcc'] = _clean_features(mfcc(waveform, n_mfcc=EXTRACTOR_ARGS['n_mfcc'], **FEATURE_ARGS))
return row_dict
def CalcFrameSpectrogram(self, plot=False, spec_mode='hz'):
fft_size = 2048
hop_length = 512 #fft_size//8
self.CurrentSpectrogram = [] #initialize for every new frame!!!
if self.CurrentFrame.ndim == 1:
#reshape so that it has dimension (dim,1)
self.CurrentFrame = np.reshape(self.CurrentFrame,
(len(self.CurrentFrame), 1))
mels_keep = 119 #128 is equal to keeping all the coefficients
for ch in range(self.CurrentFrame.shape[1]): #for every channel
Spectrogram = melspectrogram(self.CurrentFrame[:, ch],
sr=self.Fs,
n_fft=fft_size,
hop_length=hop_length,
power=2).astype(
np.float32)[0:mels_keep, :]
self.CurrentSpectrogram.append(Spectrogram)
if plot == True:
plt.figure(figsize=(10, 4))
librosa.display.specshow(librosa.power_to_db(Spectrogram,
ref=np.max),
sr=self.Fs,
hop_length=hop_length,
x_axis='time',
y_axis=spec_mode,
fmax=self.Fs / 2,
cmap='RdBu')
#plt.ylim([0, freqs[117]])
plt.colorbar(format='%+2.0f dB')
plt.title('Spectrogram')
plt.tight_layout()
plt.show()
def result():
if request.method == 'POST':
testing = []
f = request.files['file']
f.save(f.filename)
file = f.filename
model = load_model('Birdmodel.h5')
Y, SR = load(file, res_type='kaiser_fast')
onsets = frames_to_time(onset_detect(Y, SR), SR)
leading_silence = onsets[0]
y, sr = load(file,
offset=leading_silence,
res_type='kaiser_fast',
duration=10)
spec = melspectrogram(y=y, sr=sr)
testing.append(spec)
testing = np.asarray(testing)
result = Model.predict(model, testing)
if (np.where(result[0] == max(result[0]))[0] == 0):
spe = 'Vanellus indicus'
elif (np.where(result[0] == max(result[0]))[0] == 1):
spe = 'Acridotheres tristis'
elif (np.where(result[0] == max(result[0]))[0] == 2):
spe = 'Columba Livia'
elif (np.where(result[0] == max(result[0]))[0] == 3):
spe = 'Amaurornis phoenicurus'
elif (np.where(result[0] == max(result[0]))[0] == 4):
spe = 'Centropus sinensis'
os.remove(file)
clear_session()
return render_template("result.html", name=f.filename, species=spe)
def fbank(path, fft_span, hop_span, n_mels, fmin, fmax,affichage=False):
"""
:param path: emplacement du fichier
:param fft_span: taille de la fenetre pour la transformee de fourrier en seconde
:param hop_span: pas entre deux echantillons en seconde
:param n_mels: nombre de bandes de frequences mel
:param fmin: frequence minimale de la decomposition
:param fmax: frequence maximale de la decomposition
:param affichage: True si on veut afficher le spectrogramme
:return: Renvoie les vecteurs fbank representant le signal
X matrice representant la decomposition fbank au cours du temps (une ligne = une decomposition pour une periode hop_span, de taille n_mels)
"""
# 1ere facon d ouvrir un fichier
# wav_signal = scipy.io.wavfile.read(path)
# wav = np.array(wav_signal[1])
# s_rate = wav_signal[0]
# Deuxieme facon d ouvrir un fichier
wav, s_rate = librosa.load(path)
X = feature.melspectrogram(util.normalize(wav), s_rate, S=None, n_fft=int(np.floor(fft_span * s_rate)),
hop_length=int(np.floor(hop_span * s_rate)), n_mels=n_mels, fmin=fmin, fmax=fmax)
# #Verification nombre d'echantillons (un toutes les 10ms)
# size = X.shape
# print 'Taille de la matrice de sortie',size
# print 'Taille d un morceau de signal de 10ms que l on obtient' ,len(wav)/size[1]
# print 'taille theorique d un morceau de signal',0.01*s_rate
# print 's_rate',s_rate
# print 'longueur',wav.shape
# print wav.shape[0]/s_rate
X = np.log(X)
if affichage:
afficherSpec(X,s_rate,hop_span)
return np.transpose(X)
def make_blocking_data(self):
xData, yData = list(), list()
path = self.featurePath + self.name + '/'
for j, filename in enumerate(os.listdir(path)):
print(f"{self.name} {filename} ({j + 1})")
WavPath = path + filename
y, sr = load(WavPath, mono=True)
S = melspectrogram(y, sr).T
S = S[:-1 * (S.shape[0] % 128)]
num_chunk = S.shape[0] / 128
data_chunks = np.split(S, num_chunk)
xChunks, yChunks = list(), list()
for unit in data_chunks:
xChunks.append(unit)
yChunks.append(self.labelDict[self.name])
xData.append(xChunks)
yData.append(yChunks)
xData = [unit for record in xData for unit in record]
yData = [unit for record in yData for unit in record]
self.features = torch.tensor(data=xData, device=device)
self.labels = torch.tensor(data=yData, device=device)
print(self.features.shape)
print(self.labels.shape)
self.x_cat_data.append(self.features)
self.y_cat_data.append(self.labels)
return
def calculate_spectrograms(audio_dir, out_dir, file_type='.mp3'):
files = glob.glob(os.path.join(audio_dir, '*' + file_type))
num_files = len(files)
print(f'{num_files} audio files found')
if not os.path.exists(out_dir):
os.mkdir(out_dir)
for i, file_name in enumerate(sorted(files)):
start_time = time.time()
track_name = os.path.basename(file_name)
track_id = os.path.splitext(track_name)[0]
try:
song_name = track_to_song[track_id]
except KeyError:
continue
if song_name in wmf_item2i.keys():
audio_file = os.path.join(audio_dir,
track_name)
out_file = os.path.join(out_dir, track_id) + '.npy'
if not os.path.exists(out_file):
y, sr = load(audio_file)
mel_spectrogram = melspectrogram(y=y, sr=sr, n_fft=1024, hop_length=512, n_mels=128)
wmf_item2i = pickle.load(open('../../index_dicts.pkl', 'rb'))['item2i']
track_to_song = pickle.load(open('../../track_to_song.pkl', 'rb'))
calculate_spectrograms(audio_dir='../../data/MillionSongSubset/audio', out_dir='../../data/MillionSongSubset/spectrograms')
def transform_audio(self, y):
'''Compute the Mel spectrogram
Parameters
----------
y : np.ndarray
The audio buffer
Returns
-------
data : dict
data['mag'] : np.ndarray, shape=(n_frames, n_mels)
The Mel spectrogram
'''
n_frames = self.n_frames(get_duration(y=y, sr=self.sr))
mel = np.sqrt(melspectrogram(y=y, sr=self.sr,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmax=self.fmax)).astype(np.float32)
mel = fix_length(mel, n_frames)
if self.log:
mel = amplitude_to_db(mel, ref=np.max)
return {'mag': mel.T[self.idx]}
def get_seq_size(self, frames, sr):
"""
Get audio sequence size of audio time series when converted to mfcc-features or mel spectrogram
:param frames: audio time series
:param sr: sampling rate of frames
:return: sequence size of mfcc-converted audio
"""
if self.type == 'mfcc':
mfcc_frames = mfcc(frames,
sr,
n_fft=self.frame_length,
hop_length=self.hop_length,
n_mfcc=self.mfcc_features,
n_mels=self.n_mels)
return mfcc_frames.shape[1]
elif self.type == 'spectrogram':
spectrogram = melspectrogram(frames,
sr,
n_fft=self.frame_length,
hop_length=self.hop_length,
n_mels=self.n_mels)
return spectrogram.shape[1]
else:
raise ValueError('Not a valid feature type: ', self.type)
def transform_audio(self, y):
'''Compute the PCEN of the (log-) Mel Spectrogram
Parameters
----------
y : np.ndarray
The audio buffer
Returns
-------
data : dict
data['mag'] : np.ndarray, shape = (n_frames, n_bins)
The PCEN magnitude
'''
n_frames = self.n_frames(get_duration(y=y, sr=self.sr))
S = melspectrogram(y=y,
sr=self.sr,
hop_length=self.hop_length,
n_mels=self.n_mels,
n_fft=self.n_fft)
if self.log:
S = amplitude_to_db(S, ref=np.max)
P = pcen(S, sr=sr, hop_length=self.hop_length)
P = to_dtype(P, self.dtype)
return {
'mag': P[self.idx]
} #copied from mel spectrogram pump feature extractor
def from_audio(
cls,
audio,
n_fft=1024,
n_mels=128,
window="flattop",
win_length=256,
hop_length=32,
htk=True,
fmin=None,
fmax=None,
):
""" Create a MelSpectrogram object from an Audio object
The kwargs are cherry-picked from:
- https://librosa.org/doc/latest/generated/librosa.feature.melspectrogram.html#librosa.feature.melspectrogram
- https://librosa.org/doc/latest/generated/librosa.filters.mel.html?librosa.filters.mel
Args:
n_fft: Length of the FFT window [default: 1024]
n_mels: Number of mel bands to generate [default: 128]
window: The windowing function to use [default: "flattop"]
win_length: Each frame of audio is windowed by `window`. The window
will be of length `win_length` and then padded with zeros to match
`n_fft` [default: 256]
hop_length: Number of samples between successive frames [default: 32]
htk: use HTK formula instead of Slaney [default: True]
fmin: lowest frequency (in Hz) [default: None]
fmax: highest frequency (in Hz). If None, use `fmax = sr / 2.0` [default: None]
Returns:
opensoundscape.melspectrogram.MelSpectrogram object
"""
if not isinstance(audio, Audio):
raise TypeError("Class method expects Audio class as input")
process_fmin = fmin if fmin else 0
process_fmax = fmax if fmax else audio.sample_rate / 2
S = melspectrogram(
y=audio.samples,
sr=audio.sample_rate,
n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
window=window,
n_mels=n_mels,
htk=htk,
fmin=process_fmin,
fmax=process_fmax,
)
# Make spectrogram "right-side up"
S = S[::-1]
return cls(S, audio.sample_rate, hop_length, process_fmin,
process_fmax)
def make_stacked_mels(mono_signal,n_fft,samprate,hop_length,fmin,fmax,n_mels):
# create 3 mel spectrograms with different fft window size, all other variables the same
mel = [melspectrogram(mono_signal, sr=samprate, hop_length=hop_length, n_fft=j, n_mels=n_mels, fmin=fmin, fmax=fmax) for j in n_fft]
# turn spectrograms into log values
mel_db = [librosa.power_to_db(mel[k],ref=np.max) for k in range(len(mel))]
# re-stack these spectrograms into a single array
mel_db = np.stack(mel_db,axis=-1)
return mel_db
def melspectro(audio, sr=8000, win_len=256, n_mels=29):
hop_len = win_len // 2
return lf.melspectrogram(audio,
sr=sr,
n_fft=win_len,
hop_length=hop_len,
win_length=win_len,
n_mels=n_mels)
def mel_from_files(file_loc, MEL_PARAMS):
phonemes = dict()
mels = dict()
for i, loc in file_loc.items():
fs, phonemes[i] = wavfile.read(loc)
a_mel = melspectrogram(phonemes[i], sr=fs, **MEL_PARAMS)
mels[i] = a_mel
return mels
def mfcc(path):
# Let's make and display a mel-scaled power (energy-squared) spectrogram
# We use a small hop length of 64 here so that the
# frames line up with the beat tracker example below.
y, sr = load_files(path)
print 'claculating mfcc ' + path
S = feature.melspectrogram(y, sr=sr, n_fft=2048, hop_length=64, n_mels=128)
# Convert to log scale (dB). We'll use the peak power as reference.
log_S = logamplitude(S, ref_power=np.max)
mfcc_v = feature.mfcc(S=log_S, n_mfcc=14)
return np.sum(mfcc_v, axis=1)/mfcc_v.shape[1]
def extract_feature_spectrogram( data ):
filename, lbl = data
try:
signal,sr = librosa.load(filename)
if len(signal) < 5*sr:
print "Too short"
return filename, None, None
else:
print "OK", len(signal) / float(sr)
if sr != 22050:
print "Non standart sr", sr
return filename, None, None
signal = signal[:5*sr]
spectrogram = melspectrogram(y=signal, sr=sr, n_fft=fft_points,hop_length = fft_overlap,
fmax=5000, n_mels = mfcc_coefficients)
#spectrogram = spectrogram / spectrogram.max()
#print gmm.means_.shape
#result_features = np.vstack( [ gmm. ] )
return filename, lbl, spectrogram
except Exception,e:
print e
return filename, None, None
def mfcc(data, sr=22050, n_mfcc=20, **kwargs):
"""Mel-frequency cepstral coefficients (original function from librosa,
modified to accept additional parameters).
:parameters:
- data : np.ndarray or None
audio time series
- sr : int > 0
sampling rate of y
- n_mfcc: int
number of MFCCs to return
.. note::
additional keyword arguments are passed to the melspectrogram function.
:returns:
- M : np.ndarray, shape=(n_mfcc, S.shape[1])
MFCC sequence
"""
S = logamplitude(melspectrogram(y=data, sr=sr, **kwargs))
return np.dot(dct(n_mfcc, S.shape[0]), S)
def get_spectrogram(path):
"""Строим спектограмму из wav файла"""
y, sr = load(path)
S = melspectrogram(y, sr=sr, n_mels=100)
log_S = logamplitude(S, ref_power=np.max)
return log_S