def fileChromagramWrapper(wav_file):
if not os.path.isfile(wav_file):
raise Exception("Input audio file not found!")
[fs, x] = audioBasicIO.read_audio_file(wav_file)
x = audioBasicIO.stereo_to_mono(x)
specgram, TimeAxis, FreqAxis = sF.chromagram(x, fs, round(fs * 0.040),
round(fs * 0.040), True)
def extract_extract_audioAnalysis(audio_file, chuncksize=1):
[Fs, x] = audioBasicIO.read_audio_file(audio_file)
x = audioBasicIO.stereo_to_mono(x)
overlap = chuncksize * Fs
F, f_names = ShortTermFeatures.feature_extraction(x, Fs, Fs, overlap) # takes approx. 2.5 mins to comple
# return Zero Crossing Rate, Spectral Centroid, Spectral Spread, Spectral Entropy, Spectral Flux, Spectral Rolloff
return F[0], F[3], F[4], F[5], F[6], F[7]
def file_regression(input_file, model_name, model_type):
# Load classifier:
if not os.path.isfile(input_file):
print("fileClassification: wav file not found!")
return -1, -1, -1
regression_models = glob.glob(model_name + "_*")
regression_models2 = []
for r in regression_models:
if r[-5::] != "MEANS":
regression_models2.append(r)
regression_models = regression_models2
regression_names = []
for r in regression_models:
regression_names.append(r[r.rfind("_") + 1::])
# FEATURE EXTRACTION
# LOAD ONLY THE FIRST MODEL (for mt_win, etc)
if model_type == 'svm' or model_type == "svm_rbf" or \
model_type == 'randomforest':
_, _, _, mid_window, mid_step, short_window, short_step, compute_beat \
= load_model(regression_models[0], True)
# read audio file and convert to mono
samping_rate, signal = audioBasicIO.read_audio_file(input_file)
signal = audioBasicIO.stereo_to_mono(signal)
# feature extraction:
mid_features, s, _ = \
aF.mid_feature_extraction(signal, samping_rate, mid_window * samping_rate,
mid_step * samping_rate,
round(samping_rate * short_window),
round(samping_rate * short_step))
# long term averaging of mid-term statistics
mid_features = mid_features.mean(axis=1)
if compute_beat:
beat, beat_conf = aF.beat_extraction(s, short_step)
mid_features = np.append(mid_features, beat)
mid_features = np.append(mid_features, beat_conf)
# REGRESSION
R = []
for ir, r in enumerate(regression_models):
if not os.path.isfile(r):
print("fileClassification: input model_name not found!")
return (-1, -1, -1)
if model_type == 'svm' or model_type == "svm_rbf" \
or model_type == 'randomforest':
model, mean, std, _, _, _, _, _ = load_model(r, True)
curFV = (mid_features - mean) / std # normalization
R.append(regression_wrapper(model, model_type,
curFV)) # classification
return R, regression_names
def mid_feature_extraction_to_file(file_path,
mid_window,
mid_step,
short_window,
short_step,
output_file,
store_short_features=False,
store_csv=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 np file
"""
sampling_rate, signal = audioBasicIO.read_audio_file(file_path)
signal = audioBasicIO.stereo_to_mono(signal)
if store_short_features:
mid_features, short_features, _ = \
mid_feature_extraction(signal, sampling_rate,
round(sampling_rate * mid_window),
round(sampling_rate * mid_step),
round(sampling_rate * short_window),
(sampling_rate * short_step))
# save st features to np file
np.save(output_file + "_st", short_features)
if plot:
print("Short-term np file: " + output_file + "_st.npy saved")
if store_csv:
# store st features to CSV file
np.savetxt(output_file + "_st.csv",
short_features.T,
delimiter=",")
if plot:
print("Short-term CSV file: " + output_file + "_st.csv saved")
else:
mid_features, _, _ = \
mid_feature_extraction(signal, sampling_rate,
round(sampling_rate * mid_window),
round(sampling_rate * mid_step),
round(sampling_rate * short_window),
round(sampling_rate * short_step))
# save mt features to np file
np.save(output_file, mid_features)
if plot:
print("Mid-term np file: " + output_file + ".npy saved")
if store_csv:
np.savetxt(output_file + ".csv", mid_features.T, delimiter=",")
if plot:
print("Mid-term CSV file: " + output_file + ".csv saved")
def directory_feature_extraction_no_avg(folder_path, mid_window, mid_step,
short_window, short_step):
"""
This function extracts the mid-term features of the WAVE
files of a particular folder without averaging each file.
ARGUMENTS:
- folder_path: the path of the WAVE directory
- mid_window, mid_step: mid-term window and step (in seconds)
- short_window, short_step: short-term window and step (in seconds)
RETURNS:
- X: A feature matrix
- Y: A matrix of file labels
- filenames:
"""
wav_file_list = []
signal_idx = np.array([])
mid_features = np.array([])
types = ('*.wav', '*.aif', '*.aiff', '*.ogg')
for files in types:
wav_file_list.extend(glob.glob(os.path.join(folder_path, files)))
wav_file_list = sorted(wav_file_list)
for i, file_path in enumerate(wav_file_list):
sampling_rate, signal = audioBasicIO.read_audio_file(file_path)
if sampling_rate == 0:
continue
signal = audioBasicIO.stereo_to_mono(signal)
mid_feature_vector, _, _ = \
mid_feature_extraction(signal, sampling_rate,
round(mid_window * sampling_rate),
round(mid_step * sampling_rate),
round(sampling_rate * short_window),
round(sampling_rate * short_step))
mid_feature_vector = np.transpose(mid_feature_vector)
if len(mid_features) == 0: # append feature vector
mid_features = mid_feature_vector
signal_idx = np.zeros((mid_feature_vector.shape[0], ))
else:
mid_features = np.vstack((mid_features, mid_feature_vector))
signal_idx = np.append(
signal_idx, i * np.ones((mid_feature_vector.shape[0], )))
return mid_features, signal_idx, wav_file_list
def file_classification(input_file, model_name, model_type):
# Load classifier:
if not os.path.isfile(model_name):
print("fileClassification: input model_name not found!")
return -1, -1, -1
if not os.path.isfile(input_file):
print("fileClassification: wav file not found!")
return -1, -1, -1
if model_type == 'knn':
classifier, mean, std, classes, mid_window, mid_step, short_window, \
short_step, compute_beat = load_model_knn(model_name)
else:
classifier, mean, std, classes, mid_window, mid_step, short_window, \
short_step, compute_beat = load_model(model_name)
# read audio file and convert to mono
sampling_rate, signal = audioBasicIO.read_audio_file(input_file)
signal = audioBasicIO.stereo_to_mono(signal)
if sampling_rate == 0:
# audio file IO problem
return -1, -1, -1
if signal.shape[0] / float(sampling_rate) <= mid_window:
return -1, -1, -1
# feature extraction:
mid_features, s, _ = \
aF.mid_feature_extraction(signal, sampling_rate,
mid_window * sampling_rate,
mid_step * sampling_rate,
round(sampling_rate * short_window),
round(sampling_rate * short_step))
# long term averaging of mid-term statistics
mid_features = mid_features.mean(axis=1)
if compute_beat:
beat, beat_conf = aF.beat_extraction(s, short_step)
mid_features = np.append(mid_features, beat)
mid_features = np.append(mid_features, beat_conf)
feature_vector = (mid_features - mean) / std # normalization
# classification
class_id, probability = classifier_wrapper(classifier, model_type,
feature_vector)
return class_id, probability, classes
def directory_feature_extraction(folder_path,
mid_window,
mid_step,
short_window,
short_step,
compute_beat=True):
"""
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:
- folder_path: the path of the WAVE directory
- mid_window, mid_step: mid-term window and step (in seconds)
- short_window, short_step: short-term window and step (in seconds)
"""
mid_term_features = np.array([])
process_times = []
types = ('*.wav', '*.aif', '*.aiff', '*.mp3', '*.au', '*.ogg')
wav_file_list = []
for files in types:
wav_file_list.extend(glob.glob(os.path.join(folder_path, files)))
wav_file_list = sorted(wav_file_list)
wav_file_list2, mid_feature_names = [], []
for i, file_path in enumerate(wav_file_list):
print("Analyzing file {0:d} of {1:d}: {2:s}".format(
i + 1, len(wav_file_list), file_path))
if os.stat(file_path).st_size == 0:
print(" (EMPTY FILE -- SKIPPING)")
continue
sampling_rate, signal = audioBasicIO.read_audio_file(file_path)
if sampling_rate == 0:
continue
t1 = time.clock()
signal = audioBasicIO.stereo_to_mono(signal)
if signal.shape[0] < float(sampling_rate) / 5:
print(" (AUDIO FILE TOO SMALL - SKIPPING)")
continue
wav_file_list2.append(file_path)
if compute_beat:
mid_features, short_features, mid_feature_names = \
mid_feature_extraction(signal, sampling_rate,
round(mid_window * sampling_rate),
round(mid_step * sampling_rate),
round(sampling_rate * short_window),
round(sampling_rate * short_step))
beat, beat_conf = beat_extraction(short_features, short_step)
else:
mid_features, _, mid_feature_names = \
mid_feature_extraction(signal, sampling_rate,
round(mid_window * sampling_rate),
round(mid_step * sampling_rate),
round(sampling_rate * short_window),
round(sampling_rate * short_step))
mid_features = np.transpose(mid_features)
mid_features = mid_features.mean(axis=0)
# long term averaging of mid-term statistics
if (not np.isnan(mid_features).any()) and \
(not np.isinf(mid_features).any()):
if compute_beat:
mid_features = np.append(mid_features, beat)
mid_features = np.append(mid_features, beat_conf)
if len(mid_term_features) == 0:
# append feature vector
mid_term_features = mid_features
else:
mid_term_features = np.vstack(
(mid_term_features, mid_features))
t2 = time.clock()
duration = float(len(signal)) / sampling_rate
process_times.append((t2 - t1) / duration)
if len(process_times) > 0:
print("Feature extraction complexity ratio: "
"{0:.1f} x realtime".format(
(1.0 / np.mean(np.array(process_times)))))
return mid_term_features, wav_file_list2, mid_feature_names
def speaker_diarization(filename, n_speakers, mid_window=2.0, mid_step=0.2,
short_window=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)
- mid_window (opt) mid-term window size
- mid_step (opt) mid-term window step
- short_window (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
"""
sampling_rate, signal = audioBasicIO.read_audio_file(filename)
signal = audioBasicIO.stereo_to_mono(signal)
duration = len(signal) / sampling_rate
base_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)),
"data/models")
classifier_all, mean_all, std_all, class_names_all, _, _, _, _, _ = \
at.load_model_knn(os.path.join(base_dir, "knn_speaker_10"))
classifier_fm, mean_fm, std_fm, class_names_fm, _, _, _, _, _ = \
at.load_model_knn(os.path.join(base_dir, "knn_speaker_male_female"))
mid_feats, st_feats, _ = \
mtf.mid_feature_extraction(signal, sampling_rate,
mid_window * sampling_rate,
mid_step * sampling_rate,
round(sampling_rate * short_window),
round(sampling_rate * short_window * 0.5))
mid_term_features = np.zeros((mid_feats.shape[0] + len(class_names_all) +
len(class_names_fm), mid_feats.shape[1]))
for index in range(mid_feats.shape[1]):
feature_norm_all = (mid_feats[:, index] - mean_all) / std_all
feature_norm_fm = (mid_feats[:, index] - mean_fm) / std_fm
_, p1 = at.classifier_wrapper(classifier_all, "knn", feature_norm_all)
_, p2 = at.classifier_wrapper(classifier_fm, "knn", feature_norm_fm)
start = mid_feats.shape[0]
end = mid_feats.shape[0] + len(class_names_all)
mid_term_features[0:mid_feats.shape[0], index] = mid_feats[:, index]
mid_term_features[start:end, index] = p1 + 1e-4
mid_term_features[end::, index] = p2 + 1e-4
mid_feats = mid_term_features # TODO
feature_selected = [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]
mid_feats = mid_feats[feature_selected, :]
mid_feats_norm, mean, std = at.normalize_features([mid_feats.T])
mid_feats_norm = mid_feats_norm[0].T
n_wins = mid_feats.shape[1]
dist_all = np.sum(distance.squareform(distance.pdist(mid_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]
mt_feats_norm_or = mid_feats_norm
mid_feats_norm = mid_feats_norm[:, i_non_outliers]
if n_speakers <= 0:
s_range = range(2, 10)
else:
s_range = [n_speakers]
cluster_labels = []
sil_all = []
cluster_centers = []
for speakers in s_range:
k_means = sklearn.cluster.KMeans(n_clusters=speakers)
k_means.fit(mid_feats_norm.T)
cls = k_means.labels_
means = k_means.cluster_centers_
cluster_labels.append(cls)
cluster_centers.append(means)
sil_1 = []; sil_2 = []
for c in range(speakers):
# 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:
mt_feats_norm_temp = mid_feats_norm[:, cls == c]
dist = distance.pdist(mt_feats_norm_temp.T)
sil_1.append(np.mean(dist)*clust_per_cent)
sil_temp = []
for c2 in range(speakers):
if c2 != c:
clust_per_cent_2 = np.nonzero(cls == c2)[0].shape[0] /\
float(len(cls))
mid_features_temp = mid_feats_norm[:, cls == c2]
dist = distance.cdist(mt_feats_norm_temp.T,
mid_features_temp.T)
sil_temp.append(np.mean(dist)*(clust_per_cent
+ clust_per_cent_2)/2.0)
sil_temp = np.array(sil_temp)
sil_2.append(min(sil_temp))
sil_1 = np.array(sil_1)
sil_2 = np.array(sil_2)
sil = []
for c in range(speakers):
sil.append((sil_2[c] - sil_1[c]) / (max(sil_2[c], sil_1[c]) + 1e-5))
sil_all.append(np.mean(sil))
imax = int(np.argmax(sil_all))
num_speakers = s_range[imax]
cls = np.zeros((n_wins,))
for index in range(n_wins):
j = np.argmin(np.abs(index-i_non_outliers))
cls[index] = cluster_labels[imax][j]
for index in range(1):
start_prob, transmat, means, cov = \
train_hmm_compute_statistics(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)
class_names = ["speaker{0:d}".format(c) for c in range(num_speakers)]
if plot_res:
fig = plt.figure(figsize=(10, 4))
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)
list_labels = np.array(range(len(cls))) * mid_step + mid_step
ax1.set_xticks(list_labels[::25])
ax1.plot(np.array(range(len(cls))) * mid_step + mid_step, cls)
if plot_res:
plt.xlabel("time (seconds)")
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.savefig('foo.png')
return cls, sampling_rate, len(signal)
def speaker_diarization(filename, n_speakers, mid_window=2.0, mid_step=0.2,
short_window=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)
- mid_window (opt) mid-term window size
- mid_step (opt) mid-term window step
- short_window (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
"""
sampling_rate, signal = audioBasicIO.read_audio_file(filename)
signal = audioBasicIO.stereo_to_mono(signal)
duration = len(signal) / sampling_rate
base_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)),
"data/models")
classifier_all, mean_all, std_all, class_names_all, _, _, _, _, _ = \
at.load_model_knn(os.path.join(base_dir, "knn_speaker_10"))
classifier_fm, mean_fm, std_fm, class_names_fm, _, _, _, _, _ = \
at.load_model_knn(os.path.join(base_dir, "knn_speaker_male_female"))
mid_feats, st_feats, _ = \
mtf.mid_feature_extraction(signal, sampling_rate,
mid_window * sampling_rate,
mid_step * sampling_rate,
round(sampling_rate * short_window),
round(sampling_rate * short_window * 0.5))
mid_term_features = np.zeros((mid_feats.shape[0] + len(class_names_all) +
len(class_names_fm), mid_feats.shape[1]))
for index in range(mid_feats.shape[1]):
feature_norm_all = (mid_feats[:, index] - mean_all) / std_all
feature_norm_fm = (mid_feats[:, index] - mean_fm) / std_fm
_, p1 = at.classifier_wrapper(classifier_all, "knn", feature_norm_all)
_, p2 = at.classifier_wrapper(classifier_fm, "knn", feature_norm_fm)
start = mid_feats.shape[0]
end = mid_feats.shape[0] + len(class_names_all)
mid_term_features[0:mid_feats.shape[0], index] = mid_feats[:, index]
mid_term_features[start:end, index] = p1 + 1e-4
mid_term_features[end::, index] = p2 + 1e-4
mid_feats = mid_term_features # TODO
feature_selected = [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]
mid_feats = mid_feats[feature_selected, :]
mid_feats_norm, mean, std = at.normalize_features([mid_feats.T])
mid_feats_norm = mid_feats_norm[0].T
n_wins = mid_feats.shape[1]
# remove outliers:
dist_all = np.sum(distance.squareform(distance.pdist(mid_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
mt_feats_norm_or = mid_feats_norm
mid_feats_norm = mid_feats_norm[:, i_non_outliers]
# LDA dimensionality reduction:
if lda_dim > 0:
# extract mid-term features with minimum step:
window_ratio = int(round(mid_window / short_window))
step_ratio = int(round(short_window / short_window))
mt_feats_to_red = []
num_of_features = len(st_feats)
num_of_stats = 2
for index in range(num_of_stats * num_of_features):
mt_feats_to_red.append([])
# for each of the short-term features:
for index in range(num_of_features):
cur_pos = 0
feat_len = len(st_feats[index])
while cur_pos < feat_len:
n1 = cur_pos
n2 = cur_pos + window_ratio
if n2 > feat_len:
n2 = feat_len
short_features = st_feats[index][n1:n2]
mt_feats_to_red[index].append(np.mean(short_features))
mt_feats_to_red[index + num_of_features].\
append(np.std(short_features))
cur_pos += 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(class_names_all) +
len(class_names_fm),
mt_feats_to_red.shape[1]))
limit = mt_feats_to_red.shape[0] + len(class_names_all)
for index in range(mt_feats_to_red.shape[1]):
feature_norm_all = (mt_feats_to_red[:, index] - mean_all) / std_all
feature_norm_fm = (mt_feats_to_red[:, index] - mean_fm) / std_fm
_, p1 = at.classifier_wrapper(classifier_all, "knn",
feature_norm_all)
_, p2 = at.classifier_wrapper(classifier_fm, "knn", feature_norm_fm)
mt_feats_to_red_2[0:mt_feats_to_red.shape[0], index] = \
mt_feats_to_red[:, index]
mt_feats_to_red_2[mt_feats_to_red.shape[0]:limit, index] = p1 + 1e-4
mt_feats_to_red_2[limit::, index] = p2 + 1e-4
mt_feats_to_red = mt_feats_to_red_2
mt_feats_to_red = mt_feats_to_red[feature_selected, :]
mt_feats_to_red, mean, std = at.normalize_features([mt_feats_to_red.T])
mt_feats_to_red = mt_feats_to_red[0].T
labels = np.zeros((mt_feats_to_red.shape[1], ))
lda_step = 1.0
lda_step_ratio = lda_step / short_window
for index in range(labels.shape[0]):
labels[index] = int(index * short_window / lda_step_ratio)
clf = sklearn.discriminant_analysis.\
LinearDiscriminantAnalysis(n_components=lda_dim)
clf.fit(mt_feats_to_red.T, labels)
mid_feats_norm = (clf.transform(mid_feats_norm.T)).T
if n_speakers <= 0:
s_range = range(2, 10)
else:
s_range = [n_speakers]
cluster_labels = []
sil_all = []
cluster_centers = []
for speakers in s_range:
k_means = sklearn.cluster.KMeans(n_clusters=speakers)
k_means.fit(mid_feats_norm.T)
cls = k_means.labels_
means = k_means.cluster_centers_
cluster_labels.append(cls)
cluster_centers.append(means)
sil_1 = []; sil_2 = []
for c in range(speakers):
# 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 = mid_feats_norm[:, cls == c]
# compute average distance between samples
# that belong to the cluster (a values)
dist = distance.pdist(mt_feats_norm_temp.T)
sil_1.append(np.mean(dist)*clust_per_cent)
sil_temp = []
for c2 in range(speakers):
# compute distances from samples of other clusters
if c2 != c:
clust_per_cent_2 = np.nonzero(cls == c2)[0].shape[0] /\
float(len(cls))
mid_features_temp = mid_feats_norm[:, cls == c2]
dist = distance.cdist(mt_feats_norm_temp.T,
mid_features_temp.T)
sil_temp.append(np.mean(dist)*(clust_per_cent
+ clust_per_cent_2)/2.0)
sil_temp = np.array(sil_temp)
# ... and keep the minimum value (i.e.
# the distance from the "nearest" cluster)
sil_2.append(min(sil_temp))
sil_1 = np.array(sil_1)
sil_2 = np.array(sil_2)
sil = []
for c in range(speakers):
# for each cluster (speaker) compute silhouette
sil.append((sil_2[c] - sil_1[c]) / (max(sil_2[c], sil_1[c]) + 1e-5))
# keep the AVERAGE SILLOUETTE
sil_all.append(np.mean(sil))
imax = int(np.argmax(sil_all))
# optimal number of clusters
num_speakers = 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 index in range(n_wins):
j = np.argmin(np.abs(index-i_non_outliers))
cls[index] = cluster_labels[imax][j]
# Post-process method 1: hmm smoothing
for index in range(1):
# hmm training
start_prob, transmat, means, cov = \
train_hmm_compute_statistics(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)
class_names = ["speaker{0:d}".format(c) for c in range(num_speakers)]
# load ground-truth if available
gt_file = filename.replace('.wav', '.segments')
# if groundtruth exists
if os.path.isfile(gt_file):
seg_start, seg_end, seg_labs = read_segmentation_gt(gt_file)
flags_gt, class_names_gt = segments_to_labels(seg_start, seg_end,
seg_labs, mid_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))) * mid_step + mid_step / 2.0, cls)
if os.path.isfile(gt_file):
if plot_res:
ax1.plot(np.array(range(len(flags_gt))) *
mid_step + mid_step / 2.0, flags_gt, 'r')
purity_cluster_m, purity_speaker_m = \
evaluate_speaker_diarization(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)")
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 (np.array(range(len(cls))) * mid_step + mid_step / 2.0,cls)
def silence_removal(signal, sampling_rate, st_win, st_step, smooth_window=0.5,
weight=0.5, plot=False):
"""
Event Detection (silence removal)
ARGUMENTS:
- signal: the input audio signal
- sampling_rate: sampling freq
- st_win, st_step: 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:
- seg_limits: 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
signal = audioBasicIO.stereo_to_mono(signal)
st_feats, _ = stf.feature_extraction(signal, sampling_rate,
st_win * sampling_rate,
st_step * sampling_rate)
# Step 2: train binary svm classifier of low vs high energy frames
# keep only the energy short-term sequence (2nd feature)
st_energy = st_feats[1, :]
en = np.sort(st_energy)
# number of 10% of the total short-term windows
st_windows_fraction = int(len(en) / 10)
# compute "lower" 10% energy threshold
low_threshold = np.mean(en[0:st_windows_fraction]) + 1e-15
# compute "higher" 10% energy threshold
high_threshold = np.mean(en[-st_windows_fraction:-1]) + 1e-15
# get all features that correspond to low energy
low_energy = st_feats[:, np.where(st_energy <= low_threshold)[0]]
# get all features that correspond to high energy
high_energy = st_feats[:, np.where(st_energy >= high_threshold)[0]]
# form the binary classification task and ...
features = [low_energy.T, high_energy.T]
# normalize and train the respective svm probabilistic model
# (ONSET vs SILENCE)
features_norm, mean, std = at.normalize_features(features)
svm = at.train_svm(features_norm, 1.0)
# Step 3: compute onset probability based on the trained svm
prob_on_set = []
for index in range(st_feats.shape[1]):
# for each frame
cur_fv = (st_feats[:, index] - mean) / std
# get svm probability (that it belongs to the ONSET class)
prob_on_set.append(svm.predict_proba(cur_fv.reshape(1, -1))[0][1])
prob_on_set = np.array(prob_on_set)
# smooth probability:
prob_on_set = smooth_moving_avg(prob_on_set, smooth_window / st_step)
# Step 4A: detect onset frame indices:
prog_on_set_sort = np.sort(prob_on_set)
# find probability Threshold as a weighted average
# of top 10% and lower 10% of the values
nt = int(prog_on_set_sort.shape[0] / 10)
threshold = (np.mean((1 - weight) * prog_on_set_sort[0:nt]) +
weight * np.mean(prog_on_set_sort[-nt::]))
max_indices = np.where(prob_on_set > threshold)[0]
# get the indices of the frames that satisfy the thresholding
index = 0
seg_limits = []
time_clusters = []
# Step 4B: group frame indices to onset segments
while index < len(max_indices):
# for each of the detected onset indices
cur_cluster = [max_indices[index]]
if index == len(max_indices)-1:
break
while max_indices[index+1] - cur_cluster[-1] <= 2:
cur_cluster.append(max_indices[index+1])
index += 1
if index == len(max_indices)-1:
break
index += 1
time_clusters.append(cur_cluster)
seg_limits.append([cur_cluster[0] * st_step,
cur_cluster[-1] * st_step])
# Step 5: Post process: remove very small segments:
min_duration = 0.2
seg_limits_2 = []
for s_lim in seg_limits:
if s_lim[1] - s_lim[0] > min_duration:
seg_limits_2.append(s_lim)
seg_limits = seg_limits_2
if plot:
time_x = np.arange(0, signal.shape[0] / float(sampling_rate), 1.0 /
sampling_rate)
plt.subplot(2, 1, 1)
plt.plot(time_x, signal)
for s_lim in seg_limits:
plt.axvline(x=s_lim[0], color='red')
plt.axvline(x=s_lim[1], color='red')
plt.subplot(2, 1, 2)
plt.plot(np.arange(0, prob_on_set.shape[0] * st_step, st_step),
prob_on_set)
plt.title('Signal')
for s_lim in seg_limits:
plt.axvline(x=s_lim[0], color='red')
plt.axvline(x=s_lim[1], color='red')
plt.title('svm Probability')
plt.show()
return seg_limits
def mid_term_file_classification(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
"""
labels = []
accuracy = 0.0
class_names = []
cm = np.array([])
if not os.path.isfile(model_name):
print("mtFileClassificationError: input model_type not found!")
return labels, class_names, accuracy, cm
# Load classifier:
if model_type == "knn":
classifier, mean, std, class_names, mt_win, mid_step, st_win, \
st_step, compute_beat = at.load_model_knn(model_name)
else:
classifier, mean, std, class_names, mt_win, mid_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 labels, class_names, accuracy, cm
# load input file
sampling_rate, signal = audioBasicIO.read_audio_file(input_file)
# could not read file
if sampling_rate == 0:
return labels, class_names, accuracy, cm
# convert stereo (if) to mono
signal = audioBasicIO.stereo_to_mono(signal)
# mid-term feature extraction:
mt_feats, _, _ = \
mtf.mid_feature_extraction(signal, sampling_rate,
mt_win * sampling_rate,
mid_step * sampling_rate,
round(sampling_rate * st_win),
round(sampling_rate * st_step))
posterior_matrix = []
# for each feature vector (i.e. for each fix-sized segment):
for col_index in range(mt_feats.shape[1]):
# normalize current feature v
feature_vector = (mt_feats[:, col_index] - mean) / std
# classify vector:
label_predicted, posterior = \
at.classifier_wrapper(classifier, model_type, feature_vector)
labels.append(label_predicted)
# update probability matrix
posterior_matrix.append(np.max(posterior))
labels = np.array(labels)
# convert fix-sized flags to segments and classes
segs, classes = labels_to_segments(labels, mid_step)
segs[-1] = len(signal) / float(sampling_rate)
# Load grount-truth:
labels_gt, class_names_gt, accuracy, cm = \
load_ground_truth(gt_file, labels, class_names, mid_step, plot_results)
return labels, class_names, accuracy, cm
def music_thumbnailing(signal, sampling_rate, short_window=1.0, short_step=0.5,
thumb_size=10.0, limit_1=0, limit_2=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-similarity 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:
- signal: input signal
- sampling_rate: sampling frequency
- short_window: window size (in seconds)
- short_step: window step (in seconds)
- thumb_size: 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(input_file)
[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.
"""
signal = audioBasicIO.stereo_to_mono(signal)
# feature extraction:
st_feats, _ = stf.feature_extraction(signal, sampling_rate,
sampling_rate * short_window,
sampling_rate * short_step)
# self-similarity matrix
sim_matrix = self_similarity_matrix(st_feats)
# moving filter:
m_filter = int(round(thumb_size / short_step))
diagonal = np.eye(m_filter, m_filter)
sim_matrix = scipy.signal.convolve2d(sim_matrix, diagonal, 'valid')
# post-processing (remove main diagonal elements)
min_sm = np.min(sim_matrix)
for i in range(sim_matrix.shape[0]):
for j in range(sim_matrix.shape[1]):
if abs(i-j) < 5.0 / short_step or i > j:
sim_matrix[i, j] = min_sm
# find max position:
sim_matrix[0:int(limit_1 * sim_matrix.shape[0]), :] = min_sm
sim_matrix[:, 0:int(limit_1 * sim_matrix.shape[0])] = min_sm
sim_matrix[int(limit_2 * sim_matrix.shape[0])::, :] = min_sm
sim_matrix[:, int(limit_2 * sim_matrix.shape[0])::] = min_sm
rows, cols = np.unravel_index(sim_matrix.argmax(), sim_matrix.shape)
i1 = rows
i2 = rows
j1 = cols
j2 = cols
while i2-i1 < m_filter:
if i1 <= 0 or j1 <= 0 or i2 >= sim_matrix.shape[0]-2 or \
j2 >= sim_matrix.shape[1]-2:
break
if sim_matrix[i1-1, j1-1] > sim_matrix[i2 + 1, j2 + 1]:
i1 -= 1
j1 -= 1
else:
i2 += 1
j2 += 1
return short_step * i1, short_step * i2, short_step * j1, short_step * j2, \
sim_matrix
