def mtFeatureExtractionToFile(fileName, midTermSize, midTermStep, shortTermSize, shortTermStep, outPutFile,
storeStFeatures=False, storeToCSV=False, PLOT=False):
"""
This function is used as a wrapper to:
a) read the content of a WAV file
b) perform mid-term feature extraction on that signal
c) write the mid-term feature sequences to a numpy file
"""
[Fs, x] = audio_basic_io.read_audio_file(fileName) # read the wav file
x = audio_basic_io.stereo2mono(x) # convert to MONO if required
if storeStFeatures:
[mtF, stF] = mt_feature_extraction(x, Fs, round(Fs * midTermSize), round(Fs * midTermStep),
round(Fs * shortTermSize), round(Fs * shortTermStep))
else:
[mtF, _] = mt_feature_extraction(x, Fs, round(Fs * midTermSize), round(Fs * midTermStep),
round(Fs * shortTermSize), round(Fs * shortTermStep))
numpy.save(outPutFile, mtF) # save mt features to numpy file
if PLOT:
print("Mid-term numpy file: " + outPutFile + ".npy saved")
if storeToCSV:
numpy.savetxt(outPutFile + ".csv", mtF.T, delimiter=",")
if PLOT:
print("Mid-term CSV file: " + outPutFile + ".csv saved")
if storeStFeatures:
numpy.save(outPutFile + "_st", stF) # save st features to numpy file
if PLOT:
print("Short-term numpy file: " + outPutFile + "_st.npy saved")
if storeToCSV:
numpy.savetxt(outPutFile + "_st.csv", stF.T, delimiter=",") # store st features to CSV file
if PLOT:
print("Short-term CSV file: " + outPutFile + "_st.csv saved")
def fileChromagramWrapper(wavFileName):
if not os.path.isfile(wavFileName):
raise Exception("Input audio file not found!")
[Fs, x] = audio_basic_io.read_audio_file(wavFileName)
x = audio_basic_io.stereo2mono(x)
specgram, TimeAxis, FreqAxis = aF.stChromagram(x, Fs, round(Fs * 0.040),
round(Fs * 0.040), True)
def plot_spectorgram():
import audioFeatureExtraction as aF
Fs, x = audio_basic_io.read_audio_file(example_file)
x = audio_basic_io.stereo2mono(x)
specgram, TimeAxis, FreqAxis = aF.stSpectogram(x, Fs,
round(Fs * 0.040),
round(Fs * 0.040),
True)
def fileRegression(inputFile, modelName, modelType):
# Load classifier:
if not os.path.isfile(inputFile):
print("fileClassification: wav file not found!")
return (-1, -1, -1)
regressionModels = glob.glob(modelName + "_*")
regressionModels2 = []
for r in regressionModels:
if r[-5::] != "MEANS":
regressionModels2.append(r)
regressionModels = regressionModels2
regressionNames = []
for r in regressionModels:
regressionNames.append(r[r.rfind("_") + 1::])
# FEATURE EXTRACTION
# LOAD ONLY THE FIRST MODEL (for mtWin, etc)
if modelType == 'svm' or modelType == "svm_rbf":
[_, _, _, mtWin, mtStep, stWin, stStep,
computeBEAT] = loadSVModel(regressionModels[0], True)
elif modelType == 'randomforest':
[_, _, _, mtWin, mtStep, stWin, stStep,
computeBEAT] = loadRandomForestModel(regressionModels[0], True)
[Fs, x] = audio_basic_io.read_audio_file(
inputFile) # read audio file and convert to mono
x = audio_basic_io.stereo2mono(x)
# feature extraction:
[MidTermFeatures, s] = aF.mt_feature_extraction(x, Fs, mtWin * Fs,
mtStep * Fs,
round(Fs * stWin),
round(Fs * stStep))
MidTermFeatures = MidTermFeatures.mean(
axis=1) # long term averaging of mid-term statistics
if computeBEAT:
[beat, beatConf] = aF.beatExtraction(s, stStep)
MidTermFeatures = numpy.append(MidTermFeatures, beat)
MidTermFeatures = numpy.append(MidTermFeatures, beatConf)
# REGRESSION
R = []
for ir, r in enumerate(regressionModels):
if not os.path.isfile(r):
print("fileClassification: input modelName not found!")
return (-1, -1, -1)
if modelType == 'svm' or modelType == "svm_rbf":
[Model, MEAN, STD, mtWin, mtStep, stWin, stStep,
computeBEAT] = loadSVModel(r, True)
elif modelType == 'randomforest':
[Model, MEAN, STD, mtWin, mtStep, stWin, stStep,
computeBEAT] = loadRandomForestModel(r, True)
curFV = (MidTermFeatures - MEAN) / STD # normalization
R.append(regressionWrapper(Model, modelType, curFV)) # classification
return R, regressionNames
def dirWavFeatureExtractionNoAveraging(dirName, mtWin, mtStep, stWin, stStep):
"""
This function extracts the mid-term features of the WAVE files of a particular folder without averaging each file.
ARGUMENTS:
- dirName: the path of the WAVE directory
- mtWin, mtStep: mid-term window and step (in seconds)
- stWin, stStep: short-term window and step (in seconds)
RETURNS:
- X: A feature matrix
- Y: A matrix of file labels
- filenames:
"""
allMtFeatures = numpy.array([])
signalIndices = numpy.array([])
processingTimes = []
types = ('*.wav', '*.aif', '*.aiff')
wavFilesList = []
for files in types:
wavFilesList.extend(glob.glob(os.path.join(dirName, files)))
wavFilesList = sorted(wavFilesList)
for i, wavFile in enumerate(wavFilesList):
[Fs, x] = audio_basic_io.read_audio_file(wavFile) # read file
if isinstance(x, int):
continue
x = audio_basic_io.stereo2mono(x) # convert stereo to mono
[MidTermFeatures, _] = mt_feature_extraction(x, Fs, round(mtWin * Fs), round(mtStep * Fs), round(Fs * stWin),
round(Fs * stStep)) # mid-term feature
MidTermFeatures = numpy.transpose(MidTermFeatures)
# MidTermFeatures = MidTermFeatures.mean(axis=0) # long term averaging of mid-term statistics
if len(allMtFeatures) == 0: # append feature vector
allMtFeatures = MidTermFeatures
signalIndices = numpy.zeros((MidTermFeatures.shape[0],))
else:
allMtFeatures = numpy.vstack((allMtFeatures, MidTermFeatures))
signalIndices = numpy.append(signalIndices, i * numpy.ones((MidTermFeatures.shape[0],)))
return (allMtFeatures, signalIndices, wavFilesList)
def musicThumbnailing(x, Fs, shortTermSize=1.0, shortTermStep=0.5, thumbnailSize=10.0, Limit1=0, Limit2=1):
'''
This function detects instances of the most representative part of a music recording, also called "music thumbnails".
A technique similar to the one proposed in [1], however a wider set of audio features is used instead of chroma features.
In particular the following steps are followed:
- Extract short-term audio features. Typical short-term window size: 1 second
- Compute the self-silimarity matrix, i.e. all pairwise similarities between feature vectors
- Apply a diagonal mask is as a moving average filter on the values of the self-similarty matrix.
The size of the mask is equal to the desirable thumbnail length.
- Find the position of the maximum value of the new (filtered) self-similarity matrix.
The audio segments that correspond to the diagonial around that position are the selected thumbnails
ARGUMENTS:
- x: input signal
- Fs: sampling frequency
- shortTermSize: window size (in seconds)
- shortTermStep: window step (in seconds)
- thumbnailSize: desider thumbnail size (in seconds)
RETURNS:
- A1: beginning of 1st thumbnail (in seconds)
- A2: ending of 1st thumbnail (in seconds)
- B1: beginning of 2nd thumbnail (in seconds)
- B2: ending of 2nd thumbnail (in seconds)
USAGE EXAMPLE:
import audioFeatureExtraction as aF
[Fs, x] = basicIO.readAudioFile(inputFile)
[A1, A2, B1, B2] = musicThumbnailing(x, Fs)
[1] Bartsch, M. A., & Wakefield, G. H. (2005). Audio thumbnailing of popular music using chroma-based representations.
Multimedia, IEEE Transactions on, 7(1), 96-104.
'''
x = audio_basic_io.stereo2mono(x);
# feature extraction:
stFeatures = aF.st_feature_extraction(x, Fs, Fs * shortTermSize, Fs * shortTermStep)
# self-similarity matrix
S = selfSimilarityMatrix(stFeatures)
# moving filter:
M = int(round(thumbnailSize / shortTermStep))
B = np.eye(M, M)
S = scipy.signal.convolve2d(S, B, 'valid')
# post-processing (remove main diagonal elements)
MIN = np.min(S)
for i in range(S.shape[0]):
for j in range(S.shape[1]):
if abs(i - j) < 5.0 / shortTermStep or i > j:
S[i, j] = MIN;
# find max position:
S[0:int(Limit1 * S.shape[0]), :] = MIN
S[:, 0:int(Limit1 * S.shape[0])] = MIN
S[int(Limit2 * S.shape[0])::, :] = MIN
S[:, int(Limit2 * S.shape[0])::] = MIN
maxVal = np.max(S)
[I, J] = np.unravel_index(S.argmax(), S.shape)
# plt.imshow(S)
# plt.show()
# expand:
i1 = I;
i2 = I
j1 = J;
j2 = J
while i2 - i1 < M:
if i1 <= 0 or j1 <= 0 or i2 >= S.shape[0] - 2 or j2 >= S.shape[1] - 2:
break
if S[i1 - 1, j1 - 1] > S[i2 + 1, j2 + 1]:
i1 -= 1
j1 -= 1
else:
i2 += 1
j2 += 1
return (shortTermStep * i1, shortTermStep * i2, shortTermStep * j1, shortTermStep * j2, S)
def speaker_diarization(file_name, num_speaker, mt_size=2.0,
mt_step=0.2, st_win=0.05, st_step=0.025,
lda_dim=35,
plot=False):
'''
ARGUMENTS:
- fileName: the name of the WAV file to be analyzed
- numOfSpeakers the number of speakers (clusters) in the recording (<=0 for unknown)
- mtSize (opt) mid-term window size
- mtStep (opt) mid-term window step
- stWin (opt) short-term window size
- LDAdim (opt) LDA dimension (0 for no LDA)
- PLOT (opt) 0 for not plotting the results 1 for plottingy
'''
fr, x = audio_basic_io.read_audio_file(file_name)
x = audio_basic_io.stereo2mono(x)
duration = len(x) / fr
classifier1, mean1, std1, class_names1, mt_win1, mt_step1, st_win1, st_step1, compute_beta1 = aT.loadKNNModel(
os.path.join("data", "knnSpeakerAll"))
classifier2, mean2, std2, class_names2, mt_win2, mt_step2, st_win2, st_step2, compute_beta2 = aT.loadKNNModel(
os.path.join("data", "knnSpeakerFemaleMale"))
mid_term_features, short_term_features = aF.mt_feature_extraction(signal=x,
fr=fr,
mt_win=mt_size * fr,
mt_step=mt_step * fr,
st_win=round(fr * st_win),
st_step=round(fr * st_step))
# (68, 329) (34, 2630)
print(mid_term_features.shape, short_term_features.shape)
mid_term_features2 = np.zeros((mid_term_features.shape[0] + len(class_names1) + len(class_names2),
mid_term_features.shape[1]))
for i in range(mid_term_features.shape[1]):
cur_f1 = (mid_term_features[:, i] - mean1) / std1
cur_f2 = (mid_term_features[:, i] - mean2) / std2
result, p1 = aT.classifierWrapper(classifier1, "knn", cur_f1)
result, p2 = aT.classifierWrapper(classifier2, "knn", cur_f2)
mid_term_features2[0:mid_term_features.shape[0], i] = mid_term_features[:, i]
mid_term_features2[mid_term_features.shape[0]:mid_term_features.shape[0] + len(class_names1), i] = p1 + 0.0001
mid_term_features2[mid_term_features.shape[0] + len(class_names1)::, i] = p2 + 0.0001
mid_term_features = mid_term_features2 # TODO
# SELECT FEATURES:
# iFeaturesSelect = [8,9,10,11,12,13,14,15,16,17,18,19,20]; # SET 0A
# iFeaturesSelect = [8,9,10,11,12,13,14,15,16,17,18,19,20, 99,100]; # SET 0B
# iFeaturesSelect = [8,9,10,11,12,13,14,15,16,17,18,19,20, 68,69,70,71,72,73,
# 74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,
# 97,98, 99,100]; # SET 0C
i_features_select = [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 41,
42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53] # SET 1A
# iFeaturesSelect = [8,9,10,11,12,13,14,15,16,17,18,19,20,41,42,43,44,45,46,47,48,49,50,51,52,53, 99,100]; # SET 1B
# iFeaturesSelect = [8,9,10,11,12,13,14,15,16,17,18,19,20,41,42,43,44,45,46,47,
# 48,49,50,51,52,53, 68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,
# 87,88,89,90,91,92,93,94,95,96,97,98, 99,100]; # SET 1C
# iFeaturesSelect = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,34,35,
# 36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53]; # SET 2A
# iFeaturesSelect = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,34,35,
# 36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53, 99,100]; # SET 2B
# iFeaturesSelect = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,34,35,
# 36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53, 68,69,70,71,72,73,74,75,
# 76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98, 99,100]; # SET 2C
# iFeaturesSelect = range(100); # SET 3
# MidTermFeatures += np.random.rand(MidTermFeatures.shape[0], MidTermFeatures.shape[1]) * 0.000000010
mid_term_features = mid_term_features[i_features_select, :]
mid_term_features_norm, mean, std = aT.normalizeFeatures([mid_term_features.T])
mid_term_features_norm = mid_term_features_norm[0].T
num_of_windows = mid_term_features.shape[1]
# remove outliers:
distances_all = np.sum(distance.squareform(distance.pdist(mid_term_features_norm.T)), axis=0)
m_distances_all = np.mean(distances_all)
i_non_out_liers = np.nonzero(distances_all < 1.2 * m_distances_all)[0]
# TODO: Combine energy threshold for outlier removal:
# EnergyMin = np.min(MidTermFeatures[1,:])
# EnergyMean = np.mean(MidTermFeatures[1,:])
# Thres = (1.5*EnergyMin + 0.5*EnergyMean) / 2.0
# iNonOutLiers = np.nonzero(MidTermFeatures[1,:] > Thres)[0]
# print(iNonOutLiers
# per_out_lier = (100.0 * (num_of_windows - i_non_out_liers.shape[0])) / num_of_windows
mid_term_features_norm_or = mid_term_features_norm
mid_term_features_norm = mid_term_features_norm[:, i_non_out_liers]
# LDA dimensionality reduction:
if lda_dim > 0:
mt_win_ratio = int(round(mt_size / st_win))
mt_step_ratio = int(round(st_win / st_win))
mt_features_to_reduce = []
num_of_features = len(short_term_features)
num_of_statistics = 2
# for i in range(numOfStatistics * numOfFeatures + 1):
for i in range(num_of_statistics * num_of_features):
mt_features_to_reduce.append([])
for i in range(num_of_features): # for each of the short-term features:
cur_pos = 0
n = len(short_term_features[i])
while cur_pos < n:
n1 = cur_pos
n2 = cur_pos + mt_win_ratio
if n2 > n:
n2 = n
cur_st_features = short_term_features[i][n1:n2]
mt_features_to_reduce[i].append(np.mean(cur_st_features))
mt_features_to_reduce[i + num_of_features].append(np.std(cur_st_features))
cur_pos += mt_step_ratio
mt_features_to_reduce = np.array(mt_features_to_reduce)
mt_features_to_reduce2 = np.zeros((mt_features_to_reduce.shape[0] + len(class_names1) + len(class_names2),
mt_features_to_reduce.shape[1]))
for i in range(mt_features_to_reduce.shape[1]):
cur_f1 = (mt_features_to_reduce[:, i] - mean1) / std1
cur_f2 = (mt_features_to_reduce[:, i] - mean2) / std2
result, p1 = aT.classifierWrapper(classifier1, "knn", cur_f1)
result, p2 = aT.classifierWrapper(classifier2, "knn", cur_f2)
mt_features_to_reduce2[0:mt_features_to_reduce.shape[0], i] = mt_features_to_reduce[:, i]
mt_features_to_reduce2[mt_features_to_reduce.shape[0]:mt_features_to_reduce.shape[0] + len(class_names1),
i] = p1 + 0.0001
mt_features_to_reduce2[mt_features_to_reduce.shape[0] + len(class_names1)::, i] = p2 + 0.0001
mt_features_to_reduce = mt_features_to_reduce2
mt_features_to_reduce = mt_features_to_reduce[i_features_select, :]
# mtFeaturesToReduce += np.random.rand(mtFeaturesToReduce.shape[0], mtFeaturesToReduce.shape[1]) * 0.0000010
mt_features_to_reduce, mean, std = aT.normalizeFeatures([mt_features_to_reduce.T])
mt_features_to_reduce = mt_features_to_reduce[0].T
# DistancesAll = np.sum(distance.squareform(distance.pdist(mtFeaturesToReduce.T)), axis=0)
# MDistancesAll = np.mean(DistancesAll)
# iNonOutLiers2 = np.nonzero(DistancesAll < 3.0*MDistancesAll)[0]
# mtFeaturesToReduce = mtFeaturesToReduce[:, iNonOutLiers2]
labels = np.zeros((mt_features_to_reduce.shape[1],))
lda_step = 1.0
lda_step_ratio = lda_step / st_win
# print(LDAstep, LDAstepRatio
for i in range(labels.shape[0]):
labels[i] = int(i * st_win / lda_step_ratio)
clf = sklearn.discriminant_analysis.LinearDiscriminantAnalysis(n_components=lda_dim)
clf.fit(mt_features_to_reduce.T, labels)
mid_term_features_norm = (clf.transform(mid_term_features_norm.T)).T
if num_speaker <= 0:
s_range = range(2, 10)
else:
s_range = [num_speaker]
cls_all = []
sil_all = []
centers_all = []
# (26, 314)
print('mid_term_features_norm', mid_term_features_norm.shape)
for i_speakers in s_range:
k_means = sklearn.cluster.KMeans(n_clusters=i_speakers)
k_means.fit(mid_term_features_norm.T)
cls = k_means.labels_
means = k_means.cluster_centers_
# Y = distance.squareform(distance.pdist(MidTermFeaturesNorm.T))
cls_all.append(cls)
centers_all.append(means)
sil_a = []
sil_b = []
for c in range(i_speakers): # for each speaker (i.e. for each extracted cluster)
cluster_percent = np.nonzero(cls == c)[0].shape[0] / float(len(cls))
if cluster_percent < 0.020:
sil_a.append(0.0)
sil_b.append(0.0)
else:
mid_term_features_norm_temp = mid_term_features_norm[:, cls == c] # get subset of feature vectors
# compute average distance between samples that belong to the cluster (a values)
yt = distance.pdist(mid_term_features_norm_temp.T)
sil_a.append(np.mean(yt) * cluster_percent)
sil_bs = []
for c2 in range(i_speakers): # compute distances from samples of other clusters
if c2 != c:
cluster_percent2 = np.nonzero(cls == c2)[0].shape[0] / float(len(cls))
mid_term_features_norm_temp2 = mid_term_features_norm[:, cls == c2]
yt = distance.cdist(mid_term_features_norm_temp.T, mid_term_features_norm_temp2.T)
sil_bs.append(np.mean(yt) * (cluster_percent + cluster_percent2) / 2.0)
sil_bs = np.array(sil_bs)
# ... and keep the minimum value (i.e. the distance from the "nearest" cluster)
sil_b.append(min(sil_bs))
sil_a = np.array(sil_a)
sil_b = np.array(sil_b)
sil = []
for c in range(i_speakers): # for each cluster (speaker)
sil.append((sil_b[c] - sil_a[c]) / (max(sil_b[c], sil_a[c]) + 0.00001)) # compute silhouette
sil_all.append(np.mean(sil)) # keep the AVERAGE SILLOUETTE
# silAll = silAll * (1.0/(np.power(np.array(sRange),0.5)))
imax = np.argmax(sil_all) # position of the maximum sillouette value
n_speakers_final = s_range[imax] # optimal number of clusters
# generate the final set of cluster labels
# (important: need to retrieve the outlier windows:
# this is achieved by giving them the value of their nearest non-outlier window)
cls = np.zeros((num_of_windows,))
for i in range(num_of_windows):
j = np.argmin(np.abs(i - i_non_out_liers))
cls[i] = cls_all[imax][j]
# Post-process method 1: hmm smoothing
for i in range(1):
startprob, transmat, means, cov = trainHMM_computeStatistics(mid_term_features_norm_or, cls)
hmm = hmmlearn.hmm.GaussianHMM(startprob.shape[0], "diag") # hmm training
hmm.startprob_ = startprob
hmm.transmat_ = transmat
hmm.means_ = means
hmm.covars_ = cov
cls = hmm.predict(mid_term_features_norm_or.T)
# Post-process method 2: median filtering:
cls = scipy.signal.medfilt(cls, 13)
cls = scipy.signal.medfilt(cls, 11)
sil = sil_all[imax] # final sillouette
class_names = ["speaker{0:d}".format(c) for c in range(n_speakers_final)]
# load ground-truth if available
gt_file = file_name.replace('.wav', '.segments') # open for annotated file
if os.path.isfile(gt_file): # if groundturh exists
seg_start, seg_end, seg_labels = readSegmentGT(gt_file) # read GT data
flags_gt, class_names_gt = segs2flags(seg_start, seg_end, seg_labels, mt_step) # convert to flags
x = np.arange(len(cls)) * mt_step + mt_step / 2.0
if plot:
fig = plt.figure()
if num_speaker > 0:
ax1 = fig.add_subplot(111)
else:
ax1 = fig.add_subplot(211)
ax1.set_yticks(np.array(range(len(class_names))))
ax1.axis((0, duration, -1, len(class_names)))
ax1.set_yticklabels(class_names)
ax1.plot(x, cls)
if os.path.isfile(gt_file):
if plot:
ax1.plot(np.array(range(len(flags_gt))) * mt_step + mt_step / 2.0, flags_gt, 'r')
purity_cluster_mean, purity_speaker_mean = evaluateSpeakerDiarization(cls, flags_gt)
print("{0:.1f}\t{1:.1f}".format(100 * purity_cluster_mean, 100 * purity_speaker_mean))
if plot:
plt.title("Cluster purity: {0:.1f}% - Speaker purity: {1:.1f}%".format(100 * purity_cluster_mean,
100 * purity_speaker_mean))
if plot:
plt.xlabel("time (seconds)")
# print(sRange, silAll)
if num_speaker <= 0:
plt.subplot(212)
plt.plot(s_range, sil_all)
plt.xlabel("number of clusters")
plt.ylabel("average clustering's sillouette")
plt.show()
return x, cls
def silenceRemoval(x, Fs, stWin, stStep, smoothWindow=0.5, Weight=0.5, plot=False):
'''
Event Detection (silence removal)
ARGUMENTS:
- x: the input audio signal
- Fs: sampling freq
- stWin, stStep: window size and step in seconds
- smoothWindow: (optinal) smooth window (in seconds)
- Weight: (optinal) weight factor (0 < Weight < 1) the higher, the more strict
- plot: (optinal) True if results are to be plotted
RETURNS:
- segmentLimits: list of segment limits in seconds (e.g [[0.1, 0.9], [1.4, 3.0]] means that
the resulting segments are (0.1 - 0.9) seconds and (1.4, 3.0) seconds
'''
if Weight >= 1:
Weight = 0.99
if Weight <= 0:
Weight = 0.01
# Step 1: feature extraction
x = audio_basic_io.stereo2mono(x) # convert to mono
ShortTermFeatures = aF.st_feature_extraction(x, Fs, stWin * Fs, stStep * Fs) # extract short-term features
# Step 2: train binary SVM classifier of low vs high energy frames
EnergySt = ShortTermFeatures[1, :] # keep only the energy short-term sequence (2nd feature)
E = np.sort(EnergySt) # sort the energy feature values:
L1 = int(len(E) / 10) # number of 10% of the total short-term windows
T1 = np.mean(E[0:L1]) + 0.000000000000001 # compute "lower" 10% energy threshold
T2 = np.mean(E[-L1:-1]) + 0.000000000000001 # compute "higher" 10% energy threshold
Class1 = ShortTermFeatures[:, np.where(EnergySt <= T1)[0]] # get all features that correspond to low energy
Class2 = ShortTermFeatures[:, np.where(EnergySt >= T2)[0]] # get all features that correspond to high energy
featuresSS = [Class1.T, Class2.T] # form the binary classification task and ...
[featuresNormSS, MEANSS, STDSS] = aT.normalizeFeatures(featuresSS) # normalize and ...
SVM = aT.trainSVM(featuresNormSS, 1.0) # train the respective SVM probabilistic model (ONSET vs SILENCE)
# Step 3: compute onset probability based on the trained SVM
ProbOnset = []
for i in range(ShortTermFeatures.shape[1]): # for each frame
curFV = (ShortTermFeatures[:, i] - MEANSS) / STDSS # normalize feature vector
ProbOnset.append(
SVM.predict_proba(curFV.reshape(1, -1))[0][1]) # get SVM probability (that it belongs to the ONSET class)
ProbOnset = np.array(ProbOnset)
ProbOnset = smoothMovingAvg(ProbOnset, smoothWindow / stStep) # smooth probability
# Step 4A: detect onset frame indices:
ProbOnsetSorted = np.sort(
ProbOnset) # find probability Threshold as a weighted average of top 10% and lower 10% of the values
Nt = ProbOnsetSorted.shape[0] / 10
Nt = int(Nt)
T = (np.mean((1 - Weight) * ProbOnsetSorted[0:Nt]) + Weight * np.mean(ProbOnsetSorted[-Nt::]))
MaxIdx = np.where(ProbOnset > T)[0] # get the indices of the frames that satisfy the thresholding
i = 0
timeClusters = []
segmentLimits = []
# Step 4B: group frame indices to onset segments
while i < len(MaxIdx): # for each of the detected onset indices
curCluster = [MaxIdx[i]]
if i == len(MaxIdx) - 1:
break
while MaxIdx[i + 1] - curCluster[-1] <= 2:
curCluster.append(MaxIdx[i + 1])
i += 1
if i == len(MaxIdx) - 1:
break
i += 1
timeClusters.append(curCluster)
segmentLimits.append([curCluster[0] * stStep, curCluster[-1] * stStep])
# Step 5: Post process: remove very small segments:
minDuration = 0.2
segmentLimits2 = []
for s in segmentLimits:
if s[1] - s[0] > minDuration:
segmentLimits2.append(s)
segmentLimits = segmentLimits2
if plot:
timeX = np.arange(0, x.shape[0] / float(Fs), 1.0 / Fs)
plt.subplot(2, 1, 1)
plt.plot(timeX, x)
for s in segmentLimits:
plt.axvline(x=s[0])
plt.axvline(x=s[1])
plt.subplot(2, 1, 2)
plt.plot(np.arange(0, ProbOnset.shape[0] * stStep, stStep), ProbOnset)
plt.title('Signal')
for s in segmentLimits:
plt.axvline(x=s[0])
plt.axvline(x=s[1])
plt.title('SVM Probability')
plt.show()
return segmentLimits
def mtFileClassification(inputFile, modelName, modelType, plotResults=False, gtFile=""):
'''
This function performs mid-term classification of an audio stream.
Towards this end, supervised knowledge is used, i.e. a pre-trained classifier.
ARGUMENTS:
- inputFile: path of the input WAV file
- modelName: name of the classification model
- modelType: svm or knn depending on the classifier type
- plotResults: True if results are to be plotted using matplotlib along with a set of statistics
RETURNS:
- segs: a sequence of segment's endpoints: segs[i] is the endpoint of the i-th segment (in seconds)
- classes: a sequence of class flags: class[i] is the class ID of the i-th segment
'''
if not os.path.isfile(modelName):
print("mtFileClassificationError: input modelType not found!")
return (-1, -1, -1, -1)
# Load classifier:
if (modelType == 'svm') or (modelType == 'svm_rbf'):
[Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep, computeBEAT] = aT.loadSVModel(modelName)
elif modelType == 'knn':
[Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep, computeBEAT] = aT.loadKNNModel(modelName)
elif modelType == 'randomforest':
[Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep, computeBEAT] = aT.loadRandomForestModel(
modelName)
elif modelType == 'gradientboosting':
[Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep, computeBEAT] = aT.loadGradientBoostingModel(
modelName)
elif modelType == 'extratrees':
[Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep, computeBEAT] = aT.loadExtraTreesModel(
modelName)
if computeBEAT:
print("Model " + modelName + " contains long-term music features (beat etc) and cannot be used in segmentation")
return (-1, -1, -1, -1)
[Fs, x] = audio_basic_io.read_audio_file(inputFile) # load input file
if Fs == -1: # could not read file
return (-1, -1, -1, -1)
x = audio_basic_io.stereo2mono(x) # convert stereo (if) to mono
Duration = len(x) / Fs
# mid-term feature extraction:
[MidTermFeatures, _] = aF.mt_feature_extraction(x, Fs, mtWin * Fs, mtStep * Fs, round(Fs * stWin),
round(Fs * stStep))
flags = []
Ps = []
flagsInd = []
for i in range(MidTermFeatures.shape[1]): # for each feature vector (i.e. for each fix-sized segment):
curFV = (MidTermFeatures[:, i] - MEAN) / STD # normalize current feature vector
[Result, P] = aT.classifierWrapper(Classifier, modelType, curFV) # classify vector
flagsInd.append(Result)
flags.append(classNames[int(Result)]) # update class label matrix
Ps.append(np.max(P)) # update probability matrix
flagsInd = np.array(flagsInd)
# 1-window smoothing
for i in range(1, len(flagsInd) - 1):
if flagsInd[i - 1] == flagsInd[i + 1]:
flagsInd[i] = flagsInd[i + 1]
(segs, classes) = flags2segs(flags, mtStep) # convert fix-sized flags to segments and classes
segs[-1] = len(x) / float(Fs)
# Load grount-truth:
if os.path.isfile(gtFile):
[segStartGT, segEndGT, segLabelsGT] = readSegmentGT(gtFile)
flagsGT, classNamesGT = segs2flags(segStartGT, segEndGT, segLabelsGT, mtStep)
flagsIndGT = []
for j, fl in enumerate(flagsGT): # "align" labels with GT
if classNamesGT[flagsGT[j]] in classNames:
flagsIndGT.append(classNames.index(classNamesGT[flagsGT[j]]))
else:
flagsIndGT.append(-1)
flagsIndGT = np.array(flagsIndGT)
CM = np.zeros((len(classNamesGT), len(classNamesGT)))
for i in range(min(flagsInd.shape[0], flagsIndGT.shape[0])):
CM[int(flagsIndGT[i]), int(flagsInd[i])] += 1
else:
CM = []
flagsIndGT = np.array([])
acc = plotSegmentationResults(flagsInd, flagsIndGT, classNames, mtStep, not plotResults)
if acc >= 0:
print("Overall Accuracy: {0:.3f}".format(acc))
return (flagsInd, classNamesGT, acc, CM)
else:
return (flagsInd, classNames, acc, CM)
def fileClassification(inputFile, modelName, modelType):
# Load classifier:
if not os.path.isfile(modelName):
print("fileClassification: input modelName not found!")
return (-1, -1, -1)
if not os.path.isfile(inputFile):
print("fileClassification: wav file not found!")
return (-1, -1, -1)
if (modelType) == 'svm' or (modelType == 'svm_rbf'):
[
Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep,
computeBEAT
] = loadSVModel(modelName)
elif modelType == 'knn':
[
Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep,
computeBEAT
] = loadKNNModel(modelName)
elif modelType == 'randomforest':
[
Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep,
computeBEAT
] = loadRandomForestModel(modelName)
elif modelType == 'gradientboosting':
[
Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep,
computeBEAT
] = loadGradientBoostingModel(modelName)
elif modelType == 'extratrees':
[
Classifier, MEAN, STD, classNames, mtWin, mtStep, stWin, stStep,
computeBEAT
] = loadExtraTreesModel(modelName)
[Fs, x] = audio_basic_io.read_audio_file(
inputFile) # read audio file and convert to mono
x = audio_basic_io.stereo2mono(x)
if isinstance(x, int): # audio file IO problem
return (-1, -1, -1)
if x.shape[0] / float(Fs) <= mtWin:
return (-1, -1, -1)
# feature extraction:
[MidTermFeatures, s] = aF.mt_feature_extraction(x, Fs, mtWin * Fs,
mtStep * Fs,
round(Fs * stWin),
round(Fs * stStep))
MidTermFeatures = MidTermFeatures.mean(
axis=1) # long term averaging of mid-term statistics
if computeBEAT:
[beat, beatConf] = aF.beatExtraction(s, stStep)
MidTermFeatures = numpy.append(MidTermFeatures, beat)
MidTermFeatures = numpy.append(MidTermFeatures, beatConf)
curFV = (MidTermFeatures - MEAN) / STD # normalization
[Result, P] = classifierWrapper(Classifier, modelType,
curFV) # classification
return Result, P, classNames
def dirWavFeatureExtraction(dirName, mtWin, mtStep, stWin, stStep, computeBEAT=False):
"""
This function extracts the mid-term features of the WAVE files of a particular folder.
The resulting feature vector is extracted by long-term averaging the mid-term features.
Therefore ONE FEATURE VECTOR is extracted for each WAV file.
ARGUMENTS:
- dirName: the path of the WAVE directory
- mtWin, mtStep: mid-term window and step (in seconds)
- stWin, stStep: short-term window and step (in seconds)
"""
allMtFeatures = numpy.array([])
processingTimes = []
types = ('*.wav', '*.aif', '*.aiff', '*.mp3', '*.au')
wavFilesList = []
for files in types:
wavFilesList.extend(glob.glob(os.path.join(dirName, files)))
wavFilesList = sorted(wavFilesList)
wavFilesList2 = []
for i, wavFile in enumerate(wavFilesList):
print("Analyzing file {0:d} of {1:d}: {2:s}".format(i + 1, len(wavFilesList), wavFile.encode('utf-8')))
if os.stat(wavFile).st_size == 0:
print(" (EMPTY FILE -- SKIPPING)")
continue
[Fs, x] = audio_basic_io.read_audio_file(wavFile) # read file
if isinstance(x, int):
continue
t1 = time.clock()
x = audio_basic_io.stereo2mono(x) # convert stereo to mono
if x.shape[0] < float(Fs) / 10:
print(" (AUDIO FILE TOO SMALL - SKIPPING)")
continue
wavFilesList2.append(wavFile)
if computeBEAT: # mid-term feature extraction for current file
[MidTermFeatures, stFeatures] = mt_feature_extraction(x, Fs, round(mtWin * Fs), round(mtStep * Fs),
round(Fs * stWin), round(Fs * stStep))
[beat, beatConf] = beatExtraction(stFeatures, stStep)
else:
[MidTermFeatures, _] = mt_feature_extraction(x, Fs, round(mtWin * Fs), round(mtStep * Fs),
round(Fs * stWin), round(Fs * stStep))
MidTermFeatures = numpy.transpose(MidTermFeatures)
MidTermFeatures = MidTermFeatures.mean(axis=0) # long term averaging of mid-term statistics
if (not numpy.isnan(MidTermFeatures).any()) and (not numpy.isinf(MidTermFeatures).any()):
if computeBEAT:
MidTermFeatures = numpy.append(MidTermFeatures, beat)
MidTermFeatures = numpy.append(MidTermFeatures, beatConf)
if len(allMtFeatures) == 0: # append feature vector
allMtFeatures = MidTermFeatures
else:
allMtFeatures = numpy.vstack((allMtFeatures, MidTermFeatures))
t2 = time.clock()
duration = float(len(x)) / Fs
processingTimes.append((t2 - t1) / duration)
if len(processingTimes) > 0:
print("Feature extraction complexity ratio: {0:.1f} x realtime".format(
(1.0 / numpy.mean(numpy.array(processingTimes)))))
return (allMtFeatures, wavFilesList2)