def hmmSegmentation(wav_file_name, hmm_model_name, plot_res=False,
gt_file_name=""):
[fs, x] = audioBasicIO.read_audio_file(wav_file_name)
try:
fo = open(hmm_model_name, "rb")
except IOError:
print("didn't find file")
return
try:
hmm = cPickle.load(fo)
classes_all = cPickle.load(fo)
mt_win = cPickle.load(fo)
mt_step = cPickle.load(fo)
except:
fo.close()
fo.close()
[Features, _, _] = aF.mid_term_feature_extraction(x, fs, mt_win * fs,
mt_step * fs,
round(fs * 0.050),
round(fs * 0.050))
flags_ind = hmm.predict(Features.T) # apply model
if os.path.isfile(gt_file_name):
[seg_start, seg_end, seg_labs] = readSegmentGT(gt_file_name)
flags_gt, class_names_gt = segs2flags(seg_start, seg_end, seg_labs,
mt_step)
flagsGTNew = []
for j, fl in enumerate(flags_gt):
# "align" labels with GT
if class_names_gt[flags_gt[j]] in classes_all:
flagsGTNew.append(classes_all.index(class_names_gt[
flags_gt[j]]))
else:
flagsGTNew.append(-1)
cm = np.zeros((len(classes_all), len(classes_all)))
flags_ind_gt = np.array(flagsGTNew)
for i in range(min(flags_ind.shape[0], flags_ind_gt.shape[0])):
cm[int(flags_ind_gt[i]),int(flags_ind[i])] += 1
else:
flags_ind_gt = np.array([])
acc = plotSegmentationResults(flags_ind, flags_ind_gt, classes_all,
mt_step, not plot_res)
if acc >= 0:
print("Overall Accuracy: {0:.2f}".format(acc))
return (flags_ind, class_names_gt, acc, cm)
else:
return (flags_ind, classes_all, -1, -1)
def trainHMM_fromFile(wav_file, gt_file, hmm_model_name, mt_win, mt_step):
"""
This function trains a HMM model for segmentation-classification
using a single annotated audio file
ARGUMENTS:
- wav_file: the path of the audio filename
- gt_file: the path of the ground truth filename
(a csv file of the form <segment start in seconds>,
<segment end in seconds>,<segment label> in each row
- hmm_model_name: the name of the HMM model to be stored
- mt_win: mid-term window size
- mt_step: mid-term window step
RETURNS:
- hmm: an object to the resulting HMM
- class_names: a list of class_names
After training, hmm, class_names, along with the mt_win and mt_step
values are stored in the hmm_model_name file
"""
[seg_start, seg_end, seg_labs] = readSegmentGT(gt_file)
flags, class_names = segs2flags(seg_start, seg_end, seg_labs, mt_step)
[fs, x] = audioBasicIO.read_audio_file(wav_file)
[F, _, _] = aF.mid_term_feature_extraction(x, fs, mt_win * fs, mt_step * fs,
round(fs * 0.050), round(fs * 0.050))
start_prob, transmat, means, cov = trainHMM_computeStatistics(F, flags)
hmm = hmmlearn.hmm.GaussianHMM(start_prob.shape[0], "diag")
hmm.startprob_ = start_prob
hmm.transmat_ = transmat
hmm.means_ = means
hmm.covars_ = cov
fo = open(hmm_model_name, "wb")
cPickle.dump(hmm, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(class_names, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(mt_win, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(mt_step, fo, protocol=cPickle.HIGHEST_PROTOCOL)
fo.close()
return hmm, class_names
def speakerDiarization(filename, n_speakers, mt_size=2.0, mt_step=0.2,
st_win=0.05, lda_dim=35, plot_res=False):
"""
ARGUMENTS:
- filename: the name of the WAV file to be analyzed
- n_speakers the number of speakers (clusters) in
the recording (<=0 for unknown)
- mt_size (opt) mid-term window size
- mt_step (opt) mid-term window step
- st_win (opt) short-term window size
- lda_dim (opt LDA dimension (0 for no LDA)
- plot_res (opt) 0 for not plotting the results 1 for plotting
"""
[fs, x] = audioBasicIO.read_audio_file(filename)
x = audioBasicIO.stereo_to_mono(x)
duration = len(x) / fs
[classifier_1, MEAN1, STD1, classNames1, mtWin1, mtStep1, stWin1, stStep1, computeBEAT1] = aT.load_model_knn(os.path.join(os.path.dirname(os.path.realpath(__file__)), "data", "knnSpeakerAll"))
[classifier_2, MEAN2, STD2, classNames2, mtWin2, mtStep2, stWin2, stStep2, computeBEAT2] = aT.load_model_knn(os.path.join(os.path.dirname(os.path.realpath(__file__)), "data", "knnSpeakerFemaleMale"))
[mt_feats, st_feats, _] = aF.mid_term_feature_extraction(x, fs, mt_size * fs,
mt_step * fs,
round(fs * st_win),
round(fs*st_win * 0.5))
MidTermFeatures2 = np.zeros((mt_feats.shape[0] + len(classNames1) +
len(classNames2), mt_feats.shape[1]))
for i in range(mt_feats.shape[1]):
cur_f1 = (mt_feats[:, i] - MEAN1) / STD1
cur_f2 = (mt_feats[:, i] - MEAN2) / STD2
[res, P1] = aT.classifierWrapper(classifier_1, "knn", cur_f1)
[res, P2] = aT.classifierWrapper(classifier_2, "knn", cur_f2)
MidTermFeatures2[0:mt_feats.shape[0], i] = mt_feats[:, i]
MidTermFeatures2[mt_feats.shape[0]:mt_feats.shape[0]+len(classNames1), i] = P1 + 0.0001
MidTermFeatures2[mt_feats.shape[0] + len(classNames1)::, i] = P2 + 0.0001
mt_feats = MidTermFeatures2 # TODO
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]
mt_feats = mt_feats[iFeaturesSelect, :]
(mt_feats_norm, MEAN, STD) = aT.normalizeFeatures([mt_feats.T])
mt_feats_norm = mt_feats_norm[0].T
n_wins = mt_feats.shape[1]
# remove outliers:
dist_all = np.sum(distance.squareform(distance.pdist(mt_feats_norm.T)),
axis=0)
m_dist_all = np.mean(dist_all)
i_non_outliers = np.nonzero(dist_all < 1.2 * m_dist_all)[0]
# TODO: Combine energy threshold for outlier removal:
#EnergyMin = np.min(mt_feats[1,:])
#EnergyMean = np.mean(mt_feats[1,:])
#Thres = (1.5*EnergyMin + 0.5*EnergyMean) / 2.0
#i_non_outliers = np.nonzero(mt_feats[1,:] > Thres)[0]
#print i_non_outliers
perOutLier = (100.0 * (n_wins - i_non_outliers.shape[0])) / n_wins
mt_feats_norm_or = mt_feats_norm
mt_feats_norm = mt_feats_norm[:, i_non_outliers]
# LDA dimensionality reduction:
if lda_dim > 0:
#[mt_feats_to_red, _, _] = aF.mtFeatureExtraction(x, fs, mt_size * fs,
# st_win * fs, round(fs*st_win), round(fs*st_win));
# extract mid-term features with minimum step:
mt_win_ratio = int(round(mt_size / st_win))
mt_step_ratio = int(round(st_win / st_win))
mt_feats_to_red = []
num_of_features = len(st_feats)
num_of_stats = 2
#for i in range(num_of_stats * num_of_features + 1):
for i in range(num_of_stats * num_of_features):
mt_feats_to_red.append([])
for i in range(num_of_features): # for each of the short-term features:
curPos = 0
N = len(st_feats[i])
while (curPos < N):
N1 = curPos
N2 = curPos + mt_win_ratio
if N2 > N:
N2 = N
curStFeatures = st_feats[i][N1:N2]
mt_feats_to_red[i].append(np.mean(curStFeatures))
mt_feats_to_red[i+num_of_features].append(np.std(curStFeatures))
curPos += mt_step_ratio
mt_feats_to_red = np.array(mt_feats_to_red)
mt_feats_to_red_2 = np.zeros((mt_feats_to_red.shape[0] +
len(classNames1) + len(classNames2),
mt_feats_to_red.shape[1]))
for i in range(mt_feats_to_red.shape[1]):
cur_f1 = (mt_feats_to_red[:, i] - MEAN1) / STD1
cur_f2 = (mt_feats_to_red[:, i] - MEAN2) / STD2
[res, P1] = aT.classifierWrapper(classifier_1, "knn", cur_f1)
[res, P2] = aT.classifierWrapper(classifier_2, "knn", cur_f2)
mt_feats_to_red_2[0:mt_feats_to_red.shape[0], i] = mt_feats_to_red[:, i]
mt_feats_to_red_2[mt_feats_to_red.shape[0]:mt_feats_to_red.shape[0] + len(classNames1), i] = P1 + 0.0001
mt_feats_to_red_2[mt_feats_to_red.shape[0]+len(classNames1)::, i] = P2 + 0.0001
mt_feats_to_red = mt_feats_to_red_2
mt_feats_to_red = mt_feats_to_red[iFeaturesSelect, :]
#mt_feats_to_red += np.random.rand(mt_feats_to_red.shape[0], mt_feats_to_red.shape[1]) * 0.0000010
(mt_feats_to_red, MEAN, STD) = aT.normalizeFeatures([mt_feats_to_red.T])
mt_feats_to_red = mt_feats_to_red[0].T
#dist_all = np.sum(distance.squareform(distance.pdist(mt_feats_to_red.T)), axis=0)
#m_dist_all = np.mean(dist_all)
#iNonOutLiers2 = np.nonzero(dist_all < 3.0*m_dist_all)[0]
#mt_feats_to_red = mt_feats_to_red[:, iNonOutLiers2]
Labels = np.zeros((mt_feats_to_red.shape[1], ));
LDAstep = 1.0
LDAstepRatio = LDAstep / st_win
#print LDAstep, LDAstepRatio
for i in range(Labels.shape[0]):
Labels[i] = int(i*st_win/LDAstepRatio);
clf = sklearn.discriminant_analysis.LinearDiscriminantAnalysis(n_components=lda_dim)
clf.fit(mt_feats_to_red.T, Labels)
mt_feats_norm = (clf.transform(mt_feats_norm.T)).T
if n_speakers <= 0:
s_range = range(2, 10)
else:
s_range = [n_speakers]
clsAll = []
sil_all = []
centersAll = []
for iSpeakers in s_range:
k_means = sklearn.cluster.KMeans(n_clusters=iSpeakers)
k_means.fit(mt_feats_norm.T)
cls = k_means.labels_
means = k_means.cluster_centers_
# Y = distance.squareform(distance.pdist(mt_feats_norm.T))
clsAll.append(cls)
centersAll.append(means)
sil_1 = []; sil_2 = []
for c in range(iSpeakers):
# for each speaker (i.e. for each extracted cluster)
clust_per_cent = np.nonzero(cls == c)[0].shape[0] / \
float(len(cls))
if clust_per_cent < 0.020:
sil_1.append(0.0)
sil_2.append(0.0)
else:
# get subset of feature vectors
mt_feats_norm_temp = mt_feats_norm[:, cls==c]
# compute average distance between samples
# that belong to the cluster (a values)
Yt = distance.pdist(mt_feats_norm_temp.T)
sil_1.append(np.mean(Yt)*clust_per_cent)
silBs = []
for c2 in range(iSpeakers):
# compute distances from samples of other clusters
if c2 != c:
clust_per_cent_2 = np.nonzero(cls == c2)[0].shape[0] /\
float(len(cls))
MidTermFeaturesNormTemp2 = mt_feats_norm[:, cls == c2]
Yt = distance.cdist(mt_feats_norm_temp.T,
MidTermFeaturesNormTemp2.T)
silBs.append(np.mean(Yt)*(clust_per_cent
+ clust_per_cent_2)/2.0)
silBs = np.array(silBs)
# ... and keep the minimum value (i.e.
# the distance from the "nearest" cluster)
sil_2.append(min(silBs))
sil_1 = np.array(sil_1);
sil_2 = np.array(sil_2);
sil = []
for c in range(iSpeakers):
# for each cluster (speaker) compute silhouette
sil.append( ( sil_2[c] - sil_1[c]) / (max(sil_2[c],
sil_1[c]) + 0.00001))
# keep the AVERAGE SILLOUETTE
sil_all.append(np.mean(sil))
imax = np.argmax(sil_all)
# optimal number of clusters
nSpeakersFinal = s_range[imax]
# 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((n_wins,))
for i in range(n_wins):
j = np.argmin(np.abs(i-i_non_outliers))
cls[i] = clsAll[imax][j]
# Post-process method 1: hmm smoothing
for i in range(1):
# hmm training
start_prob, transmat, means, cov = \
trainHMM_computeStatistics(mt_feats_norm_or, cls)
hmm = hmmlearn.hmm.GaussianHMM(start_prob.shape[0], "diag")
hmm.startprob_ = start_prob
hmm.transmat_ = transmat
hmm.means_ = means; hmm.covars_ = cov
cls = hmm.predict(mt_feats_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]
class_names = ["speaker{0:d}".format(c) for c in range(nSpeakersFinal)];
# load ground-truth if available
gt_file = filename.replace('.wav', '.segments')
# if groundturh exists
if os.path.isfile(gt_file):
[seg_start, seg_end, seg_labs] = readSegmentGT(gt_file)
flags_gt, class_names_gt = segs2flags(seg_start, seg_end, seg_labs, mt_step)
if plot_res:
fig = plt.figure()
if n_speakers > 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(np.array(range(len(cls)))*mt_step+mt_step/2.0, cls)
if os.path.isfile(gt_file):
if plot_res:
ax1.plot(np.array(range(len(flags_gt))) *
mt_step + mt_step / 2.0, flags_gt, 'r')
purity_cluster_m, purity_speaker_m = \
evaluateSpeakerDiarization(cls, flags_gt)
print("{0:.1f}\t{1:.1f}".format(100 * purity_cluster_m,
100 * purity_speaker_m))
if plot_res:
plt.title("Cluster purity: {0:.1f}% - "
"Speaker purity: {1:.1f}%".format(100 * purity_cluster_m,
100 * purity_speaker_m))
if plot_res:
plt.xlabel("time (seconds)")
#print s_range, sil_all
if n_speakers<=0:
plt.subplot(212)
plt.plot(s_range, sil_all)
plt.xlabel("number of clusters");
plt.ylabel("average clustering's sillouette");
plt.show()
return cls
def mtFileClassification(input_file, model_name, model_type,
plot_results=False, gt_file=""):
"""
This function performs mid-term classification of an audio stream.
Towards this end, supervised knowledge is used,
i.e. a pre-trained classifier.
ARGUMENTS:
- input_file: path of the input WAV file
- model_name: name of the classification model
- model_type: svm or knn depending on the classifier type
- plot_results: 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(model_name):
print("mtFileClassificationError: input model_type not found!")
return (-1, -1, -1, -1)
# Load classifier:
if model_type == "knn":
[classifier, MEAN, STD, class_names, mt_win, mt_step, st_win, st_step,
compute_beat] = aT.load_model_knn(model_name)
else:
[classifier, MEAN, STD, class_names, mt_win, mt_step, st_win, st_step,
compute_beat] = aT.load_model(model_name)
if compute_beat:
print("Model " + model_name + " contains long-term music features "
"(beat etc) and cannot be used in "
"segmentation")
return (-1, -1, -1, -1)
[fs, x] = audioBasicIO.read_audio_file(input_file) # load input file
if fs == -1: # could not read file
return (-1, -1, -1, -1)
x = audioBasicIO.stereo_to_mono(x) # convert stereo (if) to mono
# mid-term feature extraction:
[mt_feats, _, _] = aF.mid_term_feature_extraction(x, fs, mt_win * fs,
mt_step * fs,
round(fs * st_win),
round(fs * st_step))
flags = []
Ps = []
flags_ind = []
# for each feature vector (i.e. for each fix-sized segment):
for i in range(mt_feats.shape[1]):
cur_fv = (mt_feats[:, i] - MEAN) / STD # normalize current feature v
# classify vector:
[res, P] = aT.classifierWrapper(classifier, model_type, cur_fv)
flags_ind.append(res)
flags.append(class_names[int(res)]) # update class label matrix
Ps.append(np.max(P)) # update probability matrix
flags_ind = np.array(flags_ind)
# 1-window smoothing
for i in range(1, len(flags_ind) - 1):
if flags_ind[i-1] == flags_ind[i + 1]:
flags_ind[i] = flags_ind[i + 1]
# convert fix-sized flags to segments and classes
(segs, classes) = flags2segs(flags, mt_step)
segs[-1] = len(x) / float(fs)
# Load grount-truth:
if os.path.isfile(gt_file):
[seg_start_gt, seg_end_gt, seg_l_gt] = readSegmentGT(gt_file)
flags_gt, class_names_gt = segs2flags(seg_start_gt, seg_end_gt,
seg_l_gt, mt_step)
flags_ind_gt = []
for j, fl in enumerate(flags_gt):
# "align" labels with GT
if class_names_gt[flags_gt[j]] in class_names:
flags_ind_gt.append(class_names.index(class_names_gt[
flags_gt[j]]))
else:
flags_ind_gt.append(-1)
flags_ind_gt = np.array(flags_ind_gt)
cm = np.zeros((len(class_names_gt), len(class_names_gt)))
for i in range(min(flags_ind.shape[0], flags_ind_gt.shape[0])):
cm[int(flags_ind_gt[i]),int(flags_ind[i])] += 1
else:
cm = []
flags_ind_gt = np.array([])
acc = plotSegmentationResults(flags_ind, flags_ind_gt,
class_names, mt_step, not plot_results)
if acc >= 0:
print("Overall Accuracy: {0:.3f}".format(acc) )
return (flags_ind, class_names_gt, acc, cm)
else:
return (flags_ind, class_names, acc, cm)
def trainHMM_fromDir(dirPath, hmm_model_name, mt_win, mt_step):
"""
This function trains a HMM model for segmentation-classification using
a where WAV files and .segment (ground-truth files) are stored
ARGUMENTS:
- dirPath: the path of the data diretory
- hmm_model_name: the name of the HMM model to be stored
- mt_win: mid-term window size
- mt_step: mid-term window step
RETURNS:
- hmm: an object to the resulting HMM
- class_names: a list of class_names
After training, hmm, class_names, along with the mt_win
and mt_step values are stored in the hmm_model_name file
"""
flags_all = np.array([])
classes_all = []
for i, f in enumerate(glob.glob(dirPath + os.sep + '*.wav')):
# for each WAV file
wav_file = f
gt_file = f.replace('.wav', '.segments')
if not os.path.isfile(gt_file):
continue
[seg_start, seg_end, seg_labs] = readSegmentGT(gt_file)
flags, class_names = segs2flags(seg_start, seg_end, seg_labs, mt_step)
for c in class_names:
# update class names:
if c not in classes_all:
classes_all.append(c)
[fs, x] = audioBasicIO.read_audio_file(wav_file)
[F, _, _] = aF.mid_term_feature_extraction(x, fs, mt_win * fs,
mt_step * fs, round(fs * 0.050),
round(fs * 0.050))
lenF = F.shape[1]
lenL = len(flags)
min_sm = min(lenF, lenL)
F = F[:, 0:min_sm]
flags = flags[0:min_sm]
flagsNew = []
for j, fl in enumerate(flags): # append features and labels
flagsNew.append(classes_all.index(class_names[flags[j]]))
flags_all = np.append(flags_all, np.array(flagsNew))
if i == 0:
f_all = F
else:
f_all = np.concatenate((f_all, F), axis=1)
# compute HMM statistics
start_prob, transmat, means, cov = trainHMM_computeStatistics(f_all,
flags_all)
# train the HMM
hmm = hmmlearn.hmm.GaussianHMM(start_prob.shape[0], "diag")
hmm.startprob_ = start_prob
hmm.transmat_ = transmat
hmm.means_ = means
hmm.covars_ = cov
fo = open(hmm_model_name, "wb") # save HMM model
cPickle.dump(hmm, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(classes_all, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(mt_win, fo, protocol=cPickle.HIGHEST_PROTOCOL)
cPickle.dump(mt_step, fo, protocol=cPickle.HIGHEST_PROTOCOL)
fo.close()
return hmm, classes_all