def parseAllFeatures(self, indices, filenames):
returnList = []
returnLabels = []
tot = np.zeros(len(indices))
num = 0
for el in filenames:
classname = el.split('/')[-1].strip()
# print (el, classname)
try:
[Fs, x] = audioBasicIO.readAudioFile(el)
except ValueError:
continue
F = None
if len(x.shape) == 1:
F = audioFeatureExtraction.stFeatureExtraction(
x, Fs, 0.050 * Fs, 0.025 * Fs)
else:
F = audioFeatureExtraction.stFeatureExtraction(
x[:, 0], Fs, 0.050 * Fs, 0.025 * Fs)
tot += np.mean(F[indices, :], axis=1)
num += 1
returnList.append(F[indices, :].T)
if classname[0] == 'a':
returnLabels.append(0)
elif classname[0] == 'd':
returnLabels.append(1)
elif classname[0] == 'f':
returnLabels.append(2)
elif classname[0] == 'h':
returnLabels.append(3)
elif classname[0] == 'n':
returnLabels.append(4)
elif classname[0:2] == 'sa':
returnLabels.append(5)
else:
returnLabels.append(6)
returnListLength = len(returnList)
random.seed(13921)
shuffledIndices = random.sample(range(returnListLength),
returnListLength)
shuffledReturnList = [returnList[i] for i in shuffledIndices]
shuffledReturnLabels = [returnLabels[i] for i in shuffledIndices]
return shuffledReturnList, shuffledReturnLabels
def extract_dataset(self, data, nb_samples, dataset, save=True):
f_global = []
i = 0
for (x, Fs) in data:
# 34D short-term feature
f = audioFeatureExtraction.stFeatureExtraction(x, Fs, globalvars.frame_size * Fs, globalvars.step * Fs)
# Harmonic ratio and pitch, 2D
hr_pitch = audioFeatureExtraction.stFeatureSpeed(x, Fs, globalvars.frame_size * Fs, globalvars.step * Fs)
f = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
f = stats.zscore(f, axis=0)
f = f.transpose()
f_global.append(f)
sys.stdout.write("\033[F")
i = i + 1
print("Extracting features " + str(i) + '/' + str(nb_samples) + " from data set...")
f_global = sequence.pad_sequences(f_global,
maxlen=globalvars.max_len,
dtype='float32',
padding='post',
value=globalvars.masking_value)
if save:
print("Saving features to file...")
cPickle.dump(f_global, open(dataset + '_features.p', 'wb'))
return f_global
def extract(x, sr=16000):
f_global = []
# 34D short-term feature
f = audioFeatureExtraction.stFeatureExtraction(x, sr,
globalvars.frame_size * sr,
globalvars.step * sr)
# for pyAudioAnalysis which support python3
if type(f) is tuple:
f = f[0]
# Harmonic ratio and pitch, 2D
hr_pitch = audioFeatureExtraction.stFeatureSpeed(
x, sr, globalvars.frame_size * sr, globalvars.step * sr)
f = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
f = stats.zscore(f, axis=0)
f = f.transpose()
f_global.append(f)
f_global = sequence.pad_sequences(f_global,
maxlen=globalvars.max_len,
dtype='float32',
padding='post',
value=globalvars.masking_value)
return f_global
def extract_features(path_file, frame_size=25e-3, frame_stride=10e-3):
"""Function to combine logmel and frame level ST features
extracted using pyAudioAnalysis Library
Output: 40+22 = 62 dim logmel+ST features
"""
[sample_rate, signal] = audioBasicIO.readAudioFile(path_file)
frame_length, frame_step = frame_size * sample_rate, frame_stride * sample_rate # Convert from seconds to samples
# signal_length = len(emphasized_signal)
frame_length = int(round(frame_length))
frame_step = int(round(frame_step))
st_features = audioFeatureExtraction.stFeatureExtraction(
signal, sample_rate, frame_length, frame_step)
filter_banks = extract_logmel(path_file,
frame_size=25e-3,
frame_stride=10e-3,
normalize=False)
st_features = np.transpose(
st_features) # transpose to make frame_count as x-axis
st_features = np.delete(st_features, np.s_[8:21],
axis=1) # delete the MFCCs
if st_features.shape[0] - filter_banks.shape[0] == 1:
st_features = st_features[:-1, :]
# print (st_features.shape[0], filter_banks.shape[0])
features = np.c_[st_features, filter_banks]
features -= (np.mean(features, axis=0) + 1e-8)
return features
def get_st_features(signal, rate, window_step=0.025, window_length=0.05):
"""Computes all 34 features for each window in a given signal
Parameters
----------
signal : numpy array
All sample points for the audio signal
Can be any type of number
rate : int
Sample rate of the audio signal, in Hz
window_step : float
Time step between each successive window, in seconds
Default: 0.025 (25 ms)
window_length : float
Length of each window, in seconds
Should generally be greater than windowStep to allow for overlap between frames
Default: 0.05 (50 ms)
Returns
----------
features : numpy array
NumPy array of size (number of windows) * 34
Each row in mfcc_features contains all the features for a single frame
feature_names : [str]
Names of each feature located at specified index"""
sample_step = int(rate * window_step)
sample_length = int(rate * window_length)
(features, feature_names) = audioFeatureExtraction.stFeatureExtraction(
signal, rate, sample_length, sample_step)
return features, feature_names
def load_validation_set():
"""
Output
a tuple of features: (fft features, mfcc features, mean-std features)
Description
extracts three types of features from validation set.
"""
ffts = dict()
mfccs = dict()
mean_stds = dict()
for i in validation_ids:
path = './validation/validation.{i}.wav'.format(i=i)
_, X = read_wav(path)
# FFT
fft = np.array(abs(sp.fft(X)[:1000]))
ffts.update({i: fft})
# MFCC
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
x = np.mean(ceps[int(num_ceps*1/10):int(num_ceps*9/10)], axis=0)
mfccs.update({i: x})
# Mean-Std
[Fs, x] = audioBasicIO.readAudioFile(path);
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050*Fs, 0.025*Fs);
mean_std = []
for f in F:
mean_std.extend([f.mean(), f.std()])
mean_stds.update({i: np.array(mean_std)})
return (ffts, mfccs, mean_stds)
def pitchProc(self):
print('pitchProc = ' + (self.fname))
[Fs, x] = audioBasicIO.readAudioFile(self.fname)
info = audioFeatureExtraction.stFeatureExtraction(
x, Fs, 0.050 * Fs, 0.025 * Fs)
print(len(x))
return info[0][1]
def classifyEmotion(filePath):
print("[INFO] Loading sound file")
[Fs, x] = audioBasicIO.readAudioFile(filePath)
x = audioBasicIO.stereo2mono(x)
features, _ = audioFeatureExtraction.stFeatureExtraction(
x, Fs, FRAME_SIZE * Fs, FRAME_SIZE / 2 * Fs)
inputArray = np.expand_dims(features, axis=3)
first_layer = model.get_layer(index=0)
required_input_shape = first_layer.get_config()['batch_input_shape'][1:]
# Adjust input to match required shape
if required_input_shape[1] > inputArray.shape[1]:
zerosArray = np.zeros(
(required_input_shape[0],
required_input_shape[1] - inputArray.shape[1], 1),
dtype=inputArray.dtype)
inputArray = np.concatenate((inputArray, zerosArray), axis=1)
else:
inputArray = inputArray[:, :required_input_shape[1], :]
print("[INFO] classifying sound...")
proba = model.predict(np.expand_dims(inputArray, axis=0))[0]
idx = np.argmax(proba)
label = lb.classes_[idx]
label_with_predictions = {}
for i in range(len(proba)):
label_with_predictions[lb.classes_[i]] = proba[i]
print("[INFO] Probabilities:", label_with_predictions)
print("[INFO] Prediction {}".format(label))
return label
def extract_features3(self,Fs,x):
x = audioBasicIO.stereo2mono(x) # necessary conversion for pyaudio analysis
#print len(x)
# they must be 24k samples
#coef = int(np.floor(len(x)/48000))
#x = x[range(0,len(x),6)]
#print len(x)
# Fs=16000
features = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
if len(features) == 0:
features = np.zeros((34, 2))
features_mean = np.mean(features, axis=1)
features_std = np.std(features, axis=1)
features_kurtosis = kurtosis(features, axis=1)
features_skew = skew(features, axis=1)
vec4moments = np.append(np.append(np.append(features_mean, features_std), features_kurtosis), features_skew)
result = np.asarray(vec4moments).reshape(len(vec4moments), -1).transpose()
#print(np.shape(result))
# features_complete = np.append(features_complete, features, axis=0)
return result#vec4moments # _complete
def extract_features(dataset):
data = dataset.data
nb_samples = len(dataset.targets)
frame_size = dataset.frame_size
step = dataset.step
f_global = []
i = 0
for (x, Fs) in data:
# 34D short-term feature
f = audioFeatureExtraction.stFeatureExtraction(x, Fs, frame_size * Fs, step * Fs)
# for pyAudioAnalysis which support python3
if type(f) is tuple:
f = f[0]
# Harmonic ratio and pitch, 2D
hr_pitch = audioFeatureExtraction.stFeatureSpeed(x, Fs, frame_size * Fs, step * Fs)
f = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
f = stats.zscore(f, axis=0)
f = f.transpose()
f = np.mean(f, axis=0)
f_global.append(f)
sys.stdout.write("\033[F")
i = i + 1
print("\t Extracting features " + str(i) + '/' + str(nb_samples) + " from data set...")
return f_global
def extract_features(path):
print 'extract feature of test set'
test_pkl = 'test34.pkl'
if os.path.isfile(test_pkl):
[test_set, list_filenames] = cPickle.load(open(test_pkl, 'rb'))
return test_set, list_filenames
test_set = []
list_filenames = sorted(os.listdir(path))
for filename in list_filenames:
path_to_file = os.path.join(path, filename)
[rate, sig] = audioBasicIO.readAudioFile(path_to_file)
if (rate == -1 and sig == -1):
#convert to wav
#command = "ffmpeg -i {}".format(path_to_file)
extension = os.path.splitext(filename)[-1]
new_file = path_to_file.replace(extension, '.wav')
command = "ffmpeg -i {} {}".format(path_to_file, new_file)
os.system(command)
[rate, sig] = audioBasicIO.readAudioFile(new_file)
os.system('rm {}'.format(
path_to_file)) #remove old file not in *.wav format
if sig.ndim >= 2: #merge multichannels into mono channel
sig = np.mean(sig, axis=1)
features = audioFeatureExtraction.stFeatureExtraction(
sig, rate, win * rate, step * rate)
features = features.reshape((features.shape[1], -1))
test_set.append(features)
cPickle.dump([test_set, list_filenames], open(test_pkl, 'wb'), -1)
return test_set, list_filenames
def pitchProc2(self, results_dict):
print("pitchProc2")
[Fs, x] = audioBasicIO.readAudioFile(self.fname)
info = audioFeatureExtraction.stFeatureExtraction(
x, Fs, 0.050 * Fs, 0.025 * Fs)
results_dict["pitch"] = info[0][1]
return info[0][1]
def newFeatures(inFile, isVAD):
# read audio data from file
#[Fs, x] = aIO.readAudioFile(inFile)
x, Fs = librosa.load(inFile, sr=16000) #sf.read(inFile)
# window and overlap size
win = int(0.025 * Fs)
step = int(0.010 * Fs)
# get short-time features
Feats = aFE.stFeatureExtraction(x, Fs, win, step)
if isVAD:
energy = Feats[1]
thv = energy.mean() * 0.1
i_speechs = np.where(energy > thv)[0]
Feats = Feats[:, i_speechs]
# saveFeats(Feats)
Feats = np.transpose(Feats[8:21, :])
newFeat = []
for row in range(len(Feats)):
newFeat.append(Feats[row, :])
return newFeat
def predict_genre(song_path):
optimal_training_features = load(OPTIMAL_TRAINING_FEATURES_DIR)
training_data = load(OPTIMAL_TRAINING_FEATURES_DATA_DIR).item()
X_train = training_data['X_train']
y_train = training_data['y_train']
x_test = []
Fs, x = audioBasicIO.readAudioFile(song_path)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050 * Fs, 0.025 * Fs)
features_to_extract = sorted(optimal_training_features, key=itemgetter('feature_id'))
for feature in features_to_extract:
feature_index = feature['feature_id'] - 1
feature_prop = feature['feature_prop']
feature_data = None
if feature_prop == 'min':
feature_data = min(F[feature_index, :])
elif feature_prop == 'mean':
feature_data = mean(F[feature_index, :])
elif feature_prop == 'max':
feature_data = max(F[feature_index, :])
x_test.append(feature_data)
X_train_scale, x_test_scale = scale_data_multiple(X_train, x_test)
prediction = k_nearest_neighbor.k_nearest_neighbor(X_train_scale, y_train, x_test_scale, training_data['k'])
return prediction
def getClassIDsToClipFramesMFCCs(classIDsToClipsMap):
classToMFCCsOfClips = {}
for classID in classIDsToClipsMap:
classToMFCCsOfClips[classID] = []
for singleEventAudioClip in classIDsToClipsMap[classID]:
# this returns a matrix that looks like:
"""
zcr | energy | energy_entropy | .... | mfcc_0 | mfcc_1 | ... | ... |
win0 value value
----
win1 value
----
.
.
.
"""
# so featuresMatrix[0] = a list of the zcr for each window, and featuresMatrix[0][0]
# would be the zcr for the very first window
# we care about featuresMatrix[8:20] for the 13 MFCCs
featuresMatrix, featureNames = fe.stFeatureExtraction(
singleEventAudioClip, sf, window_size, step_size)
mfccsForEachWindow = []
# for each window, loop through indicies 8-20 to get all of its mfccs into one list, and compile all of
# that into a bigger list
for i in range(0, len(featuresMatrix[8])):
mfccsForOneWindow = []
for j in range(8, 21):
mfccsForOneWindow.append(featuresMatrix[j][i])
mfccsForEachWindow.append(mfccsForOneWindow)
# have list that looks like [ [mfcc0, mfcc1, mfcc2...], [mfcc0...], ... ] for each window in one clip
classToMFCCsOfClips[classID].append(mfccsForEachWindow)
return classToMFCCsOfClips
def extractFeatures():
files = []
print "Getting example names..."
with open('./labels') as f:
for line in f:
files.append("_".join(line.split("\t")[0].split("_")[:-1]) +
'.wav')
f.close()
print "Walking through files to extract features..."
for dirpath, dirnames, filenames in os.walk(CLIPS_DIR_PATH):
for f in filenames:
if f in files:
print "Extracting for", f
path = './clips/clips/' + f
audiofile = AudioSegment.from_file(path)
data = np.fromstring(audiofile._data, np.int16)
Fs = audiofile.frame_rate
x = []
for chn in xrange(audiofile.channels):
x.append(data[chn::audiofile.channels])
x = np.array(x).T
if x.ndim == 2:
if x.shape[1] == 1:
x = x.flatten()
try:
features = audioFeatureExtraction.stFeatureExtraction(
x, Fs, 0.050 * Fs, 0.025 * Fs).T
np.save(FEATURE_PICKLES + f, features)
except ValueError as e:
print e
def process_request():
data = request.get_json(force=True)
feature_type = data['feat_type']
assert feature_type in ['raw', 'mfcc', 'all']
model_input = data['m_in']
y = pd.read_json(data['feat']).as_matrix().squeeze()
# get the sampled raw signal
if feature_type == 'raw':
sr = data['sr']
if model_input == 'mfcc':
feat = librosa.feature.mfcc(y=y, sr=int(sr),
n_mfcc=40).T # (S, 40)
feat = th.FloatTensor(feat).unsqueeze(0)
elif model_input == 'all':
all_feats, f_names = audioFeatureExtraction.stFeatureExtraction(
y, sr, 2048, 512)
feat = th.FloatTensor(all_feats.T).unsqueeze(0)
# get the mfcc directly
elif feature_type == 'mfcc':
feat = th.FloatTensor(y).view(-1, 40).unsqueeze(0)
elif feature_type == 'all':
feat = th.FloatTensor(y).view(-1, 34).unsqueeze(0)
# load model
model = SER(h_size=200, feat_size=feat.size(2), class_num=4, dropout=0.)
#model.cuda()
model.eval()
#model.load_state_dict(th.load('checkpoint/model.pt'))
pred = model(feat, [feat.size(1)], None, None)
pred = pred.max(dim=1)[1].item()
return str(pred)
def __init__(self, filename):
[Fs, x] = audioBasicIO.readAudioFile(filename)
F = audioFeatureExtraction.stFeatureExtraction(
np.mean(x, axis=1) if x.ndim == 2 else x, Fs, 0.050 * Fs,
0.025 * Fs)
# print (F[0][1])
self.input_from_audio = F[0][1]
def build_MFCC_for_one_sound_slice(folder, sound_slice):
'''
builds the MFCC coeffs, given the sound_slice and folder containing the sound_slice
:param file: str
:param sound_slice:
:return: array
folder is the folder name where sound_slice is
sound_slice must be in .wav format, not .mp3 format, in order to apply pyAusioAnalysis
'''
sound_slice_fullname = os.path.join(folder, sound_slice)
[Fs, x] = audioBasicIO.readAudioFile(sound_slice_fullname)
F, f_names = audioFeatureExtraction.stFeatureExtraction(
x, Fs, 0.050 * Fs, 0.025 * Fs)
G = resample_matrix(F,
num=15) # see the definitionog resample_matrix above
#feat_list=[]
#feat_list.append(G)
'''Below, we extract the MFCC's from the extracted and then resampled features above
.For each i, MFCC_list[i] gives for the i-th audio, all 13 feature vectors, each of them 15-dimensional
resampled values of the original feature vectors'''
MFCC_list = []
#for i in range(len(feat_list)):
#MFCC = feat_list[i][8:21, :] # 9th to 21-st features are the MFCC coeffs
MFCC = G[8:21, :] # 8th to 20 th features are the MFCC coeffs
MFCC_flat = np.ndarray.flatten(
MFCC
) # flatening the array, but are we destroying time series structure?
MFCC_flattned_as_list = list(
MFCC_flat
) #MFCC_array was np array, so we convert it into list, to avoid dim like [1, foo, bar]
#MFCC_list.append(MFCC_flat_as_list)
#MFCC_array = np.asarray(MFCC_list)
return MFCC_flattned_as_list
def extract_mfcc_features(filename):
print(filename)
[Fs, x] = audioBasicIO.readAudioFile(filename)
x = audioBasicIO.stereo2mono(x)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050 * Fs,
0.025 * Fs)
mfcc = F[8:21, :] # 13 mfcc features
return mfcc
def beatExtractionWrapper(wav_file, plot):
if not os.path.isfile(wav_file):
raise Exception("Input audio file not found!")
[fs, x] = audioBasicIO.readAudioFile(wav_file)
F, _ = aF.stFeatureExtraction(x, fs, 0.050 * fs, 0.050 * fs)
bpm, ratio = aF.beatExtraction(F, 0.050, plot)
print("Beat: {0:d} bpm ".format(int(bpm)))
print("Ratio: {0:.2f} ".format(ratio))
def beatExtractionWrapper(wav_file, plot):
if not os.path.isfile(wav_file):
raise Exception("Input audio file not found!")
[fs, x] = audioBasicIO.readAudioFile(wav_file)
F, _ = aF.stFeatureExtraction(x, fs, 0.050 * fs, 0.050 * fs)
bpm, ratio = aF.beatExtraction(F, 0.050, plot)
print("Beat: {0:d} bpm ".format(int(bpm)))
print("Ratio: {0:.2f} ".format(ratio))
def read_files(files): X = [] for fn in tqdm(files): y, sr = librosa.load(fn, sr=8000) y = preprocess(y) features = audioFeatureExtraction.stFeatureExtraction(y, sr, 0.10*sr, .05*sr) X.extend(features) return X
def featureExtractor(fileName): [Fs, x] = audioBasicIO.readAudioFile(fileName) Features = audioFeatureExtraction.stFeatureExtraction(x, Fs,0.001 * Fs, 0.0003 * Fs) MFCCs = [] for index in range(len(Features)): MFCCs.append(float(np.mean(Features[index]))) return MFCCs
def beatExtractionWrapper(wavFileName, plot):
if not os.path.isfile(wavFileName):
raise Exception("Input audio file not found!")
[Fs, x] = audioBasicIO.readAudioFile(wavFileName)
F = aF.stFeatureExtraction(x, Fs, 0.050 * Fs, 0.050 * Fs)
BPM, ratio = aF.beatExtraction(F, 0.050, plot)
print("Beat: {0:d} bpm ".format(int(BPM)))
print("Ratio: {0:.2f} ".format(ratio))
def getStVectorPerWav(wavFile, stWin,
stStep): # given a wav, get entire sT features
[Fs, x] = getTotalAudio([wavFile])
ShortTermFeatures = aF.stFeatureExtraction(x, Fs, stWin * Fs, stStep * Fs)
[featuresNormSS, MEANSS, STDSS
] = aT.normalizeFeatures([ShortTermFeatures]) # normalize to 0-mean 1-std
[X, y] = featureListToVectors([featuresNormSS])
return X, y, Fs
def _get_feature(file_path):
"""使用pyAudioAnalysis获取特征名称和对应的特征值"""
# [fs, x] = audioBasicIO.readAudioFile(file_path)
sr, x = read(file_path)
x = audioBasicIO.stereo2mono(x)
f, f_names = audioFeatureExtraction.stFeatureExtraction(x, sr, 0.050 * sr, 0.025 * sr)
return f_names, [np.mean(fm) for fm in f]
def extract_mfcc(signal: np.ndarray,
sample_rate: int = 44100,
window: float = 0.5,
stride: float = 0.25):
feats, f_names = audioFeatureExtraction.stFeatureExtraction(
signal, sample_rate, sample_rate * window, stride * sample_rate)
return feats.T, f_names
def preProcess(fileName):
[Fs, x] = audioBasicIO.readAudioFile(fileName)
if (len(x.shape) > 1 and x.shape[1] == 2):
x = np.mean(x, axis=1, keepdims=True)
else:
x = x.reshape(x.shape[0], 1)
F, f_names = audioFeatureExtraction.stFeatureExtraction(
x[:, 0], Fs, 0.050 * Fs, 0.025 * Fs)
return (f_names, F)
def get_desired_features(self, int_audio_frames):
feature_vector = fe.stFeatureExtraction(int_audio_frames,
self.sample_rate,
self.frame_size_samples,
self.frame_step_samples)
output = []
for desired_index in self.desired_features_indices:
output.append(feature_vector[desired_index])
return np.array(output)
def analyze_audio_response(filename):
#This function analyzes the yes/no response of a .wav file and returns a 34x (number of time frames depending on how long the response is) matrix
from pyAudioAnalysis import audioBasicIO
from pyAudioAnalysis import audioFeatureExtraction
import matplotlib.pyplot as plt
[Fs, x] = audioBasicIO.readAudioFile(filename)
F = audioFeatureExtraction.stFeatureExtraction(x[:, 0], Fs, Fs * 0.05,
Fs * 0.025)
return F
def extract_features(self, file_path):
[Fs, x] = audioBasicIO.readAudioFile(file_path)
x = audioBasicIO.stereo2mono(x) # necessary conversion for pyaudio analysis
features = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
features = np.mean(features, axis=1)
features = np.asarray(features).reshape(len(features), -1).transpose()
# features_complete = np.append(features_complete, features, axis=0)
return features # _complete
def process_mp3_files():
files = read_input()
os.system("touch test.wav")
for mp3_file in files:
mean_value = []
sound = AudioSegment.from_mp3(mp3_file)
sound.export("test.wav", format="wav")
# print mp3_file
[Fs, x] = audioBasicIO.readAudioFile("test.wav")
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050 * Fs, 0.025 * Fs)
for i in range(len(F)):
mean_value.append(numpy.mean(F[i]))
compute_emotion(mean_value)
def add_audio_feature_extraction(Fs, x, label_id, features, labels):
"""
Input
Fs: frequency
x: signal
label_id: label(genre) id
features: array of ffts
labels: array of labels
Description
extracts a bunch of features listed in here(https://github.com/tyiannak/pyAudioAnalysis/wiki/3.-Feature-Extraction)
from x and appends it to features.
"""
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050*Fs, 0.025*Fs);
features.append(F)
labels.append(label_id)
def showFeatures(name):
print("processing - " + name)
[Fs, x] = audioBasicIO.readAudioFile(name)
# print(x)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.50 * Fs, 0.25 * Fs)
# print(x.size, Fs, 0.50 * Fs, 0.25 * Fs)
# a = F[0, :]
# numpy.savetxt("foo.csv", a, delimiter=",")
# plt.subplot(3, 1, 1)
# plt.plot(F[0, :])
# plt.xlabel('Frame no')
# plt.ylabel('ZCR')
#
# plt.subplot(3, 1, 2)
# plt.plot(F[1, :])
# plt.xlabel('Frame no')
# plt.ylabel('Energy')
#
# plt.subplot(3, 1, 3)
# plt.plot(F[3, :])
# plt.xlabel('Frame no')
# plt.ylabel('SC')
#
# plt.show()
# items = ' '.join(map(str, a))
# print(items)
# print("--", F[0, :])
vec = [
F[0, :].mean(), F[1, :].mean(), F[4, :].mean(), F[5, :].mean(), F[6, :].mean(), F[7, :].mean(),
F[0, :].std(), F[1, :].std(), F[4, :].std(), F[5, :].std(), F[6, :].std(), F[7, :].std()
]
vecstr = ' '.join(map(str, vec))
print("vector in audio.py : ",vecstr);
melfeat = melfeature(F)
# chromafeat = chromafeature(F)
return vecstr + " " + melfeat
def main(argv):
if argv[1] == "-shortTerm":
for i in range(nExp):
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
duration = x.shape[0] / float(Fs)
t1 = time.clock()
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050*Fs, 0.050*Fs);
t2 = time.clock()
perTime1 = duration / (t2-t1); print "short-term feature extraction: {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-classifyFile":
for i in range(nExp):
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
duration = x.shape[0] / float(Fs)
t1 = time.clock()
aT.fileClassification("diarizationExample.wav", "svmSM","svm")
t2 = time.clock()
perTime1 = duration / (t2-t1); print "Mid-term feature extraction + classification \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-mtClassify":
for i in range(nExp):
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
duration = x.shape[0] / float(Fs)
t1 = time.clock()
[flagsInd, classesAll, acc] = aS.mtFileClassification("diarizationExample.wav", "svmSM", "svm", False, '')
t2 = time.clock()
perTime1 = duration / (t2-t1); print "Fix-sized classification - segmentation \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-hmmSegmentation":
for i in range(nExp):
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
duration = x.shape[0] / float(Fs)
t1 = time.clock()
aS.hmmSegmentation('diarizationExample.wav', 'hmmRadioSM', False, '')
t2 = time.clock()
perTime1 = duration / (t2-t1); print "HMM-based classification - segmentation \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-silenceRemoval":
for i in range(nExp):
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
duration = x.shape[0] / float(Fs)
t1 = time.clock()
[Fs, x] = audioBasicIO.readAudioFile("diarizationExample.wav");
segments = aS.silenceRemoval(x, Fs, 0.050, 0.050, smoothWindow = 1.0, Weight = 0.3, plot = False)
t2 = time.clock()
perTime1 = duration / (t2-t1); print "Silence removal \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-thumbnailing":
for i in range(nExp):
[Fs1, x1] = audioBasicIO.readAudioFile("scottish.wav")
duration1 = x1.shape[0] / float(Fs1)
t1 = time.clock()
[A1, A2, B1, B2, Smatrix] = aS.musicThumbnailing(x1, Fs1, 1.0, 1.0, 15.0) # find thumbnail endpoints
t2 = time.clock()
perTime1 = duration1 / (t2-t1); print "Thumbnail \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-diarization-noLDA":
for i in range(nExp):
[Fs1, x1] = audioBasicIO.readAudioFile("diarizationExample.wav")
duration1 = x1.shape[0] / float(Fs1)
t1 = time.clock()
aS.speakerDiarization("diarizationExample.wav", 4, LDAdim = 0, PLOT = False)
t2 = time.clock()
perTime1 = duration1 / (t2-t1); print "Diarization \t {0:.1f} x realtime".format(perTime1)
elif argv[1] == "-diarization-LDA":
for i in range(nExp):
[Fs1, x1] = audioBasicIO.readAudioFile("diarizationExample.wav")
duration1 = x1.shape[0] / float(Fs1)
t1 = time.clock()
aS.speakerDiarization("diarizationExample.wav", 4, PLOT = False)
t2 = time.clock()
perTime1 = duration1 / (t2-t1); print "Diarization \t {0:.1f} x realtime".format(perTime1)
raise Exception("Dimension of the covariance matrix and data should match")
invcov = cov.T
mean = np.reshape(mean, (1, n))
x = x - (np.ones((d, 1))*mean).T
fact = np.sum(((np.dot(invcov, x))*x), axis = 1)
y = np.exp(-0.5*fact)
y = np.divide(y, math.pow((2*math.pi), n)*np.std(cov))
return y
# feature extraction from the library pyAudioAnalysis
attribute = audioFeatureExtraction.stFeatureExtraction(x, Fs, SIZE_OF_WINDOW, SIZE_OF_STEP)
# relationship between the similarity and the timestamp in the audio
relation = [[1 for col in range(2)] for row in range(attribute.shape[1]/BLOCK_STEP)]
while (END_OF_FILE == 0):
if (FIRST_PAIR == 1):
# for the first pari of the block
block_i_index_start = 0
block_i_index_end = BLOCK_SIZE
block_i_attribute = getMFCCs(block_i_index_start, block_i_index_end)
block_i_mean = np.mean(block_i_attribute, axis=1)
block_i_cov = np.cov(block_i_attribute)
block_i_log_like = np.log(gauss(block_i_attribute, mean=block_i_mean, cov=block_i_cov))
def features(filename, tag): signal, sampfreq = lr.load(filename) features = afe.stFeatureExtraction(signal, sampfreq, 0.050 * sampfreq, 0.025 * sampfreq) return signal, sampfreq, features
def train_classifier():
data_set = []
for file in os.listdir("training_dataset/unhappy"):
temp = []
mean_value = []
if file.endswith(".mp3"):
# print "training_dataset/unhappy/"+file
sound = AudioSegment.from_mp3("training_dataset/unhappy/" + file)
sound.export("test.wav", format="wav")
[Fs, x] = audioBasicIO.readAudioFile("test.wav")
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
for i in range(len(F)):
temp.append(numpy.mean(F[i]))
mean_value.append(temp)
mean_value.append(1)
data_set.append(mean_value)
for file in os.listdir("training_dataset/happy"):
temp = []
mean_value = []
if file.endswith(".mp3"):
# print "training_dataset/happy/"+file
sound = AudioSegment.from_mp3("training_dataset/happy/" + file)
sound.export("test.wav", format="wav")
[Fs, x] = audioBasicIO.readAudioFile("test.wav")
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
for i in range(len(F)):
temp.append(numpy.mean(F[i]))
mean_value.append(temp)
mean_value.append(2)
data_set.append(mean_value)
for file in os.listdir("training_dataset/angry"):
temp = []
mean_value = []
if file.endswith(".mp3"):
# print "training_dataset/angry/"+file
sound = AudioSegment.from_mp3("training_dataset/angry/" + file)
sound.export("test.wav", format="wav")
[Fs, x] = audioBasicIO.readAudioFile("test.wav")
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
for i in range(len(F)):
temp.append(numpy.mean(F[i]))
mean_value.append(temp)
mean_value.append(3)
data_set.append(mean_value)
for file in os.listdir("training_dataset/neutral"):
temp = []
mean_value = []
if file.endswith(".mp3"):
# print "training_dataset/neutral/"+file
sound = AudioSegment.from_mp3("training_dataset/neutral/" + file)
sound.export("test.wav", format="wav")
[Fs, x] = audioBasicIO.readAudioFile("test.wav")
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.05 * Fs, 0.025 * Fs)
for i in range(len(F)):
temp.append(numpy.mean(F[i]))
mean_value.append(temp)
mean_value.append(4)
data_set.append(mean_value)
x = []
y = []
for i in range(len(data_set)):
x.append(data_set[i][0])
y.append(data_set[i][1])
clf = RandomForestClassifier(n_estimators=30, max_features=6, max_depth=None, min_samples_split=1, bootstrap=True)
clf = clf.fit(x, y)
f2 = open("classifier.pickle", "wb")
pickle.dump(clf, f2)
f2.close()
from pyAudioAnalysis import audioBasicIO
from pyAudioAnalysis import audioFeatureExtraction
import matplotlib.pyplot as plt
[Fs1, x1] = audioBasicIO.readAudioFile("happy.wav");
[Fs2, x2] = audioBasicIO.readAudioFile("sad.wav");
# Fs is frequency
# x is real data
th = 100 # fixed fea length
k12 = (len(x1)-800)/th/float(Fs1)
k22 = (len(x2)-800)/th/float(Fs2)
F1, f_names1 = audioFeatureExtraction.stFeatureExtraction(x1, Fs1, 0.05*Fs1, k12*Fs1);
F2, f_names2 = audioFeatureExtraction.stFeatureExtraction(x2, Fs2, 0.05*Fs2, k22*Fs2);
# stFeatureExtraction(signal, fs, win, step):
# signal: the input signal samples
# fs: the sampling freq (in Hz)
# win: the short-term window size (in samples)
# step: the short-term window step (in samples)
'''
here,
window size = 0.05*Fs = 0.05*16000 = 800
step size = 0.025*Fs = 0.024*16000 = 400
we can get n frames from signal with length 23776
400*n+800=23776 -> n=57.44 = 58
as below F.shape = (34,58)
def main(path):
ds = Dataset(path)
loader = Loader(path + "/train/", 32, 16)
X = []
y = []
Z = []
ii = 0
for p in ds.trainTracks():
f = p.split("/")
name = f[len(f) - 1]
labelTeller = loader.loadLabelsForSoundfile(name)
[Fs, x] = audioBasicIO.readAudioFile(p)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.032 * Fs, 0.016 * Fs)
G = zip(*F)
N = 0
if len(G) > labelTeller.tellNoOfAllBlocks():
N = labelTeller.tellNoOfAllBlocks()
else:
N = len(G)
for i in range(N):
Z.append([G[i], labelTeller.tell(i)])
# i = 0
# for w in ds.windows(x,44100, 1410, 705):
# mf = mfcc(w)
# row = [i]
# Z.append([mf[0],labelTeller.tell(i)])
# i = i+1
print p + " " + str(ii) + "/61"
ii = ii + 1
print "shuffle"
random.shuffle(Z)
Z = zip(*Z)
NN = 20000
L = NN
R = NN
FINAL = [[], []]
for i in range(len(Z[0])):
if Z[1][i] == "sing" and L > 0:
L = L - 1
FINAL[0].append(Z[0][i])
FINAL[1].append(Z[1][i])
if Z[1][i] == "nosing" and R > 0:
R = R - 1
FINAL[0].append(Z[0][i])
FINAL[1].append(Z[1][i])
clf = svm.SVC(cache_size=2000)
print "######### " + str(len(Z[0]))
clf.fit(FINAL[0], FINAL[1])
loader = Loader(path + "/test/", 32, 16)
print "Loading test"
for p in ds.validationTracks():
X = []
y = []
f = p.split("/")
name = f[len(f) - 1]
labelTeller = loader.loadLabelsForSoundfile(name)
i = 0
[Fs, x] = audioBasicIO.readAudioFile(p)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.032 * Fs, 0.016 * Fs)
G = zip(*F)
N = 0
if len(G) > labelTeller.tellNoOfAllBlocks():
N = labelTeller.tellNoOfAllBlocks()
else:
N = len(G)
for i in range(N):
X.append(G[i])
y.append(labelTeller.tell(i))
print "Starting prediction " + p
Y = clf.predict(X)
ok = 0
al = 0
for i in range(len(y)):
if y[i] == Y[i]:
ok = ok + 1
al = al + 1
print ok / float(al)
from pyAudioAnalysis import audioBasicIO
from pyAudioAnalysis import audioFeatureExtraction
import matplotlib.pyplot as plt
[Fs, x] = audioBasicIO.readAudioFile("../audio_data/doremi.wav")
print Fs
print len(x)
#using a frame size of 50 msecs and a frame step of 25 msecs (50% overlap)
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050*Fs, 0.025*Fs)
"""
stFeatureExtraction
This function implements the shor-term windowing process. For each short-term window a set of features is extracted.
This results to a sequence of feature vectors, stored in a numpy matrix.
ARGUMENTS
signal: the input signal samples
Fs: the sampling freq (in Hz)
Win: the short-term window size (in samples)
Step: the short-term window step (in samples)
RETURNS
stFeatures: a numpy array (numOfFeatures x numOfShortTermWindows)
"""
print len(F)
plt.subplot(2,1,1); plt.plot(F[0,:]); plt.xlabel('Frame no'); plt.ylabel('ZCR')
plt.subplot(2,1,2); plt.plot(F[1,:]); plt.xlabel('Frame no'); plt.ylabel('Energy'); plt.show()
