def test_get_annotated_data_x_y(self):
timestamps = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
data = np.random.rand(8, 10)
labels = [
AudacityLabel(1.2, 2.5, "m"),
AudacityLabel(4.5, 6.2, "s")
]
x, y, classes, timestamps = Util.get_annotated_data_x_y(timestamps, data, labels)
self.assertEqual(3, x.shape[0])
self.assertListEqual(["m", "s", "s"], y)
self.assertListEqual(["m", "s"], classes)
self.assertListEqual([0.0, 1.0, 2.0], timestamps)
labels = [
AudacityLabel(1.0, 5.5, 'A'),
AudacityLabel(5.5, 10.0, 'B'),
AudacityLabel(15.0, 20.5, 'C')
]
X, y, classes, new_timestamps = Util.get_annotated_data_x_y([float(i) for i in range(0, 25)], np.ones((25, 10)),
labels)
self.assertListEqual(['A', 'A', 'A', 'A', 'A', 'B', 'B', 'B', 'B', 'C', 'C', 'C', 'C', 'C', 'C'], y)
self.assertListEqual([0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0], new_timestamps)
def multimodal_analysis(video_path, output, tmp_dir, duration, output_path):
video_frames_dir = os.path.abspath(os.path.join(tmp_dir, 'video_frames'))
if not os.path.isdir(video_frames_dir):
extract_images_from_video(os.path.abspath(video_path),
video_frames_dir)
generate_face_based_segmentation(
output, os.path.abspath(os.path.join(tmp_dir, 'video_frames')),
os.path.abspath(output_path), args.speakers,
os.path.abspath('models/shape_predictor_68_face_landmarks.dat'),
os.path.abspath('models/dlib_face_recognition_resnet_model_v1.dat'),
tmp_dir)
# fusion
mapping_face_to_voice = calculate_fusion(
output,
os.path.abspath(output_path),
Util.read_audacity_labels(
os.path.join(os.path.abspath(output_path),
'%s.audio.txt' % output)),
Util.read_audacity_labels(
os.path.join(os.path.abspath(output_path),
'%s.image.txt' % output)),
duration,
step=0.05,
neighbours_before_after=40,
times_greater=4)
mapping_face_to_voice_json = os.path.join(
os.path.abspath(output_path), output + ".mapping_face_to_voice.json")
with open(mapping_face_to_voice_json, 'w') as f:
json.dump(mapping_face_to_voice, f)
def main():
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--input-file', dest='input_file', required=True)
args = parser.parse_args()
input_dir = os.path.split(args.input_file)[0]
temp_file = input_dir + "/temp.wav"
cmd = [
"/usr/bin/ffmpeg", "-i", args.input_file, "-ar", "22050", "-ac", "1",
"-acodec", "pcm_s16le", temp_file, "-y"
]
subprocess.check_call(cmd)
cmd = ["yaafe", "-c", FEATURE_PLAN, "-r", "22050", temp_file]
subprocess.check_output(cmd)
features1 = ["zcr", "flux", "spectral_rollof", "energy_stats"]
features2 = ["mfcc_stats"]
features3 = ["spectral_flatness_per_band"]
features4 = features1 + features2 + features3
FEATURE_GROUPS = [features1, features2, features3, features4]
peaks, convolution_values, timestamps = feat.get_combined_peaks(
temp_file, FEATURE_GROUPS, kernel_type="gaussian")
detected_segments = kernel.calculate_segment_start_end_times_from_peak_positions(
peaks, timestamps)
timestamps, feature_vectors = feat.read_features(features4,
temp_file,
scale=True)
with open("/opt/speech-music-discrimination/pickled/model.pickle",
'r') as f:
trained_model = pickle.load(f)
frame_level_predictions = trained_model.predict(feature_vectors)
annotated_segments = Util.get_annotated_labels_from_predictions_and_sm_segments(
frame_level_predictions, detected_segments, timestamps)
annotated_segments = Util.combine_adjacent_labels_of_the_same_class(
annotated_segments)
annotated_segments = feat.filter_noisy_labels(annotated_segments)
annotated_segments = Util.combine_adjacent_labels_of_the_same_class(
annotated_segments)
Util.write_audacity_labels(annotated_segments,
input_dir + "/annotated-segments.txt")
for f in glob.glob(input_dir + "/*.csv"):
os.remove(f)
os.remove(temp_file)
def combine_audio_segments(input_dir, output_name, output_dir, clear_output_dir):
if clear_output_dir:
Util.remove_dir(output_dir)
Util.make_dir(output_dir)
files_grouped_by_class = WavEditor.get_files_grouped_by_class(input_dir)
for class_name in files_grouped_by_class:
input_wavs = []
for k, file in enumerate(files_grouped_by_class[class_name]):
input_wavs.append(file)
cmd = ['sox'] + input_wavs + [os.path.join(output_dir, '%s_%s.wav' % (output_name, class_name))]
subprocess.check_call(cmd)
def main():
subprocess.check_output(["unzip", "muspeak-mirex2015-detection-examples.zip", "-d",
"muspeak-mirex2015-detection-examples"])
mp3_to_wav()
wav_files = glob.glob("./muspeak-mirex2015-detection-examples/*.wav")
for wav_file in wav_files:
print wav_file
label_file = wav_file.replace(".mp3.wav", ".csv")
if not os.path.isfile(label_file):
label_file = label_file.replace(".csv", "_v2.csv")
WavEditor.create_audio_segments(label_file, wav_file, "segments", True, ",", "f2", remove_overlapping=True)
speech_wavs = glob.glob("./segments/*_s.wav")
music_wavs = glob.glob("./segments/*_m.wav")
all_files_dict = {}
for f in speech_wavs:
all_files_dict[f] = "s"
for f in music_wavs:
all_files_dict[f] = "m"
random.seed(2222)
all_files_random_keys = random.sample(all_files_dict.keys(), len(all_files_dict.keys()))
last_seconds = 0
files_to_concatenate = []
labels = []
for v in all_files_random_keys:
duration = float(subprocess.check_output(["soxi", "-D", v]).strip())
segment_start_time = last_seconds
segment_end_time = last_seconds + duration
last_seconds += duration
labels.append(AudacityLabel(segment_start_time, segment_end_time, all_files_dict[v]))
files_to_concatenate.append(v)
audacity_labels = Util.combine_adjacent_labels_of_the_same_class(labels)
Util.write_audacity_labels(audacity_labels, "mirex_combined.txt")
command = []
command.append("sox")
command.extend(files_to_concatenate)
command.append("mirex_combined.wav")
subprocess.check_output(command)
shutil.rmtree("./segments")
shutil.rmtree("./muspeak-mirex2015-detection-examples")
def calculate_der(reference_filename, hypothesis_filename):
lbls = Util.read_audacity_labels(reference_filename)
reference = Annotation()
for lbl in lbls:
reference[Segment(lbl.start_seconds, lbl.end_seconds)] = lbl.label
predicted_lbls = Util.read_audacity_labels(hypothesis_filename)
hypothesis = Annotation()
for lbl in predicted_lbls:
if lbl.label != 'non_speech':
hypothesis[Segment(lbl.start_seconds, lbl.end_seconds)] = lbl.label
metric = DiarizationErrorRate()
der = metric(reference, hypothesis)
return der
def get_features(features, datasets_dir, pca=False):
timestamps_gtzan, feature_vectors_gtzan = Util.read_merged_features(
datasets_dir + "/gtzan/gtzan_combined.wav", features)
labels_gtzan = Util.read_audacity_labels(datasets_dir +
"/gtzan/gtzan_combined.txt")
X_gtzan, Y_gtzan, lbls_gtzan = Util.get_annotated_data_x_y(
timestamps_gtzan, feature_vectors_gtzan, labels_gtzan)
timestamps_labrosa, feature_vectors_labrosa = Util.read_merged_features(
datasets_dir + "/labrosa/labrosa_combined.wav", features)
labels_labrosa = Util.read_audacity_labels(datasets_dir +
"/labrosa/labrosa_combined.txt")
X_labrosa, Y_labrosa, lbls_labrosa = Util.get_annotated_data_x_y(
timestamps_labrosa, feature_vectors_labrosa, labels_labrosa)
timestamps_mirex, feature_vectors_mirex = Util.read_merged_features(
datasets_dir + "/mirex/mirex_combined.wav", features)
labels_mirex = Util.read_audacity_labels(datasets_dir +
"/mirex/mirex_combined.txt")
X_mirex, Y_mirex, lbls_mirex = Util.get_annotated_data_x_y(
timestamps_mirex, feature_vectors_mirex, labels_mirex)
scaler = StandardScaler()
scaler.fit(np.concatenate((X_labrosa, X_gtzan, X_mirex)))
with open("pickled/scaler.pickle", 'w') as f:
pickle.dump(scaler, f)
X_gtzan = scaler.transform(X_gtzan)
X_labrosa = scaler.transform(X_labrosa)
X_mirex = scaler.transform(X_mirex)
if pca:
pca = PCA(n_components=20)
pca.fit(np.concatenate((X_labrosa, X_gtzan, X_mirex)))
X_gtzan = pca.transform(X_gtzan)
X_labrosa = pca.transform(X_labrosa)
X_mirex = pca.transform(X_mirex)
data = {
"x_gtzan": X_gtzan,
"y_gtzan": Y_gtzan,
"labels_gtzan": labels_gtzan,
"x_labrosa": X_labrosa,
"y_labrosa": Y_labrosa,
"labels_labrosa": labels_labrosa,
"x_mirex": X_mirex,
"y_mirex": Y_mirex,
"labels_mirex": labels_mirex,
"timestamps_gtzan": timestamps_gtzan,
"timestamps_labrosa": timestamps_labrosa,
"timestamps_mirex": timestamps_mirex
}
return data
def test_non_maximum_suppression(self):
peaks = [2, 3, 4, 7, 8, 9]
values = [2, 2, 2, 3, 2, 2, 2, 4, 5, 4, 2, 2, 2]
p, v = Util.non_maximum_suppression(peaks, values)
self.assertListEqual([3, 8], p)
self.assertListEqual(values, v)
def test_get_annotation_time_shift(self):
labelsA = [
AudacityLabel(1.0, 5.0, 'A'),
AudacityLabel(5.0, 10.0, 'B'),
AudacityLabel(15.0, 20.0, 'B')
]
new_labels = Util.get_annotation_time_shift(labelsA)
expected = [
{
'old_label': AudacityLabel(1.0, 5.0, 'A'),
'new_label': AudacityLabel(0.0, 4.0, 'A'),
'shift': 1.0
},
{
'old_label': AudacityLabel(5.0, 10.0, 'B'),
'new_label': AudacityLabel(4.0, 9.0, 'B'),
'shift': 1.0
},
{
'old_label': AudacityLabel(15.0, 20.0, 'B'),
'new_label': AudacityLabel(9.0, 14.0, 'B'),
'shift': 6.0
}
]
self.assertListEqual(expected, new_labels)
def test_get_unshifted_labels(self):
predicted_lbls = [
AudacityLabel(1.0, 3.0, "A"),
AudacityLabel(8.0, 12.0, "B")
]
shifted_unshifted_labels = [
{
'old_label': AudacityLabel(1.0, 5.0, 'A'),
'new_label': AudacityLabel(0.0, 4.0, 'A'),
'shift': 1.0
},
{
'old_label': AudacityLabel(5.0, 10.0, 'B'),
'new_label': AudacityLabel(4.0, 9.0, 'B'),
'shift': 1.0
},
{
'old_label': AudacityLabel(15.0, 20.0, 'B'),
'new_label': AudacityLabel(9.0, 14.0, 'B'),
'shift': 6.0
}
]
lbls = Util.get_unshifted_labels(predicted_lbls, shifted_unshifted_labels)
self.assertListEqual([AudacityLabel(2.0, 4.0, 'A'), AudacityLabel(9.0, 18.0, 'B')], lbls)
def test_get_unshifted_timestamps(self):
lbls = [
{
'old_label': AudacityLabel(1.0, 5.0, 'A'),
'new_label': AudacityLabel(0.0, 4.0, 'A'),
'shift': 1.0
},
{
'old_label': AudacityLabel(5.0, 10.0, 'B'),
'new_label': AudacityLabel(4.0, 9.0, 'B'),
'shift': 1.0
},
{
'old_label': AudacityLabel(15.0, 20.0, 'B'),
'new_label': AudacityLabel(9.0, 14.0, 'B'),
'shift': 6.0
}
]
shifted_timestamps = [3.0, 4.0, 11.0]
expected_unshifted_timestamps = [
shifted_timestamps[0] + lbls[0]['shift'],
shifted_timestamps[1] + lbls[1]['shift'],
shifted_timestamps[2] + lbls[2]['shift']
]
unshifted_timestamps = Util.get_unshifted_timestamps(shifted_timestamps, lbls)
self.assertListEqual(expected_unshifted_timestamps, unshifted_timestamps)
def read_features(features, wavfile, scale=False):
timestamps, feature_vectors = Util.read_merged_features(wavfile, features)
if scale:
with open("/opt/speech-music-discrimination/pickled/scaler.pickle",
'r') as f:
scaler = pickle.load(f)
feature_vectors = scaler.transform(feature_vectors)
return timestamps, feature_vectors
def test_combine_peaks(self):
a_peaks = [3, 8]
a_peak_values = [2, 2, 2, 3, 2, 2, 2, 4, 5, 4, 2, 2, 2]
b_peaks = [6]
b_peak_values = [2, 2, 2, 2, 2, 2, 7, 2, 2, 2, 2, 2, 2]
p, v = Util.combine_peaks(a_peaks, a_peak_values, b_peaks, b_peak_values)
self.assertListEqual([3, 6, 8], p)
def test_get_annotated_labels_from_predictions_and_sm_segments(self):
timestamps = [0, 1, 2, 3, 4, 5, 6]
segments = [
AudacityLabel(0, 2.5, '-'),
AudacityLabel(2.5, 5.5, '-')
]
frame_level_predictions = np.array(['v', 'v', 'v', 's', 's', 'v', 'v'])
labels = Util.get_annotated_labels_from_predictions_and_sm_segments(frame_level_predictions, segments, timestamps)
self.assertListEqual(['v', 's'], [l.label for l in labels])
def get_files_grouped_by_class(input_dir):
input_dir = os.path.abspath(input_dir)
files = os.listdir(input_dir)
files_grouped_by_class = {}
for file in files:
file_class = Util.get_class_from_filename(file)
if not file_class in files_grouped_by_class:
files_grouped_by_class[file_class] = []
files_grouped_by_class[file_class].append(os.path.join(input_dir, file))
return files_grouped_by_class
def calculate_fusion(youtube_video_id,
lbls_dir,
audio_lbls,
image_lbls,
duration,
step=0.1,
neighbours_before_after=6,
times_greater=2): # 100ms
pairs, timestamps = create_pairs(audio_lbls, image_lbls, duration, step)
mapping_face_to_voice = detect_face_voice_mapping(pairs)
# print(mapping_face_to_voice)
pairs = apply_mapping_to_pairs(pairs, mapping_face_to_voice)
# print(pairs)
seconds_of_mismatch = 0
for k, pair in enumerate(pairs):
if pair.image_class is None:
# when image is None
continue
classes = pair.image_class.split(",")
# if only one face has been detected then assume it's the face of the speaker
if len(classes) == 1 and pair.audio_class != 'non_speech':
if pair.image_class != pair.audio_class:
seconds_of_mismatch += step
# print("%s != %s" % (pair.image_class, pair.audio_class))
nearest_neighbour_class = find_nearest_neighbours_class(
k,
pairs,
neighbours_before_after=neighbours_before_after,
times_greater=times_greater)
pair.audio_class = nearest_neighbour_class
lbls = Util.generate_labels_from_classifications(
[p.audio_class for p in pairs], timestamps)
lbls = list(filter(lambda x: x.label is not None, lbls))
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".fusion.txt"))
return mapping_face_to_voice
def test_split_data_based_on_annotation(self):
X = np.array([
[1, 2, 3, 4],
[2, 3, 4, 5],
[4, 5, 6, 7]
])
Y = [0, 0, 1]
classes = ["music", "speech"]
data = Util.split_data_based_on_annotation(X, Y, classes)
self.assertEqual(data["music"].shape[0], 2)
self.assertEqual(data["speech"].shape[0], 1)
def test_get_annotated_data(self):
timestamps = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
data = np.random.rand(7, 10)
labels = [
AudacityLabel(1.2, 2.5, "m"),
AudacityLabel(4.5, 6.0, "s")
]
annotated_data = Util.get_annotated_data(timestamps, data, labels)
self.assertTrue("m" in annotated_data)
self.assertTrue("s" in annotated_data)
self.assertTrue(data[2, :] in annotated_data["m"])
self.assertTrue(data[5, :] in annotated_data["s"])
self.assertTrue(data[6, :] in annotated_data["s"])
def create_audio_segments(labels, input_wav, output_dir, clear_output_dir, delimiter, format, remove_overlapping=False):
if clear_output_dir:
Util.remove_dir(output_dir)
Util.make_dir(output_dir)
rows = WavEditor.get_rows(labels, delimiter)
if format == "f2" and remove_overlapping:
rows = WavEditor.get_non_overlapping_items(rows)
for k, row in enumerate(rows):
if format == "f1" or (format == "f2" and remove_overlapping):
start_time = row[0]
end_time = row[1]
elif format == "f2":
start_time = row[0]
end_time = str(float(row[0]) + float(row[1]))
else:
logging.error("not supported file format")
sys.exit(1)
label = row[2]
filename = str(uuid.uuid4())
WavEditor.create_audio_segment(start_time, end_time, input_wav,
os.path.join(output_dir, "%s_%s.wav" % (filename, label)))
def calculate_fusion(youtube_video_id, lbls_dir, audio_lbls, image_lbls, duration, step=0.1): # 100ms
pairs, timestamps = create_pairs(audio_lbls, image_lbls, duration, step)
mapping_face_to_voice = detect_face_voice_mapping(pairs)
print(mapping_face_to_voice)
pairs = apply_mapping_to_pairs(pairs, mapping_face_to_voice)
print(pairs)
for k, pair in enumerate(pairs):
if pair.image_class is None:
# when image is None
continue
classes = pair.image_class.split(",")
if len(classes) == 1 and pair.audio_class != 'non_speech':
if pair.image_class != pair.audio_class:
# print("%s != %s" % (pair.image_class, pair.audio_class))
nearest_neighbour_class = find_nearest_neighbours_class(k, pairs)
pair.audio_class = nearest_neighbour_class
print(pairs)
lbls = Util.generate_labels_from_classifications([p.audio_class for p in pairs], timestamps)
lbls = filter(lambda x: x.label is not None, lbls)
json_lbls = []
for lbl in lbls:
json_lbls.append({
"start_seconds": lbl.start_seconds,
"end_seconds": lbl.end_seconds,
"label": lbl.label
})
with open(os.path.join(lbls_dir, youtube_video_id + ".json"), 'w') as outfile:
json.dump(json_lbls, outfile)
Util.write_audacity_labels(lbls, os.path.join(lbls_dir, youtube_video_id + ".txt"))
return mapping_face_to_voice
def test_combine_adjacent_labels_of_the_same_class(self):
input_labels = [
AudacityLabel(0, 10, "m"),
AudacityLabel(10, 20, "m"),
AudacityLabel(20, 21, "s"),
AudacityLabel(21, 22, "s"),
AudacityLabel(22, 23, "s"),
AudacityLabel(23, 30, "m")
]
expected_labels = [
AudacityLabel(0, 20, "m"),
AudacityLabel(20, 23, "s"),
AudacityLabel(23, 30, "m"),
]
actual_labels = Util.combine_adjacent_labels_of_the_same_class(input_labels)
self.assertListEqual(expected_labels, actual_labels)
def test_generate_labels_from_classifications(self):
classifications = [1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1]
timestamps = [0.0, 0.6965986394557823, 1.3931972789115645, 2.089795918367347, 2.786394557823129,
3.4829931972789114, 4.179591836734694, 4.876190476190477, 5.572789115646258, 6.2693877551020405,
6.965986394557823, 7.662585034013605, 8.359183673469389, 9.05578231292517, 9.752380952380953,
10.448979591836734, 11.145578231292516, 11.842176870748299, 12.538775510204081,
13.235374149659863, 13.931972789115646, 14.628571428571428, 15.32517006802721, 16.021768707482995]
labels = Util.generate_labels_from_classifications(classifications, timestamps)
expected_labels = [AudacityLabel(0.0, 1.3931972789115645, 1),
AudacityLabel(1.3931972789115645, 9.752380952380953, 0),
AudacityLabel(9.752380952380953, 10.448979591836736, 1),
AudacityLabel(10.448979591836734, 11.145578231292516, 0),
AudacityLabel(11.145578231292516, 12.538775510204081, 1),
AudacityLabel(12.538775510204081, 13.235374149659863, 0),
AudacityLabel(13.235374149659863, 16.718367346938777, 1)]
self.assertListEqual(expected_labels, labels)
def test_load_yaafe_csv_double_stats(self):
timestamps, features = Util.load_yaafe_csv(self.double_stats_csv)
self.assertEqual((17, 2), features.shape)
self.assertEqual(334.3673469387755, timestamps[-1])
def test_load_yaafe_csv_stats(self):
timestamps, features = Util.load_yaafe_csv(self.stats_csv)
self.assertEqual((17, 2), features.shape)
self.assertEqual(11.145578231292516, timestamps[-1])
def test_load_yaafe_csv_no_stats(self):
timestamps, features = Util.load_yaafe_csv(self.no_stats_csv)
self.assertEqual((17, 2), features.shape)
self.assertEqual(0.37151927437641724, timestamps[-1])
def test_parse_yaafe_header_stats_derivate(self):
header = Util.parse_yaafe_header(self.stats_derivate)
self.assertEqual(22050, header['samplerate'])
self.assertEqual(15360, header['effective_step_size'])
def test_calculate_classes_percentages(self):
classifications = [1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1]
percentages = Util.calculate_classes_percentages(classifications)
self.assertAlmostEqual(0.5833333333333334, percentages[0])
self.assertAlmostEqual(0.4166666666666667, percentages[1])
for f in music_wavs:
all_files_dict[f] = "m"
random.seed(1111)
all_files_random_keys = random.sample(all_files_dict.keys(), len(all_files_dict.keys()))
last_seconds = 0
files_to_concatenate = []
labels = []
for v in all_files_random_keys:
duration = float(subprocess.check_output(["soxi", "-D", v]).strip())
segment_start_time = last_seconds
segment_end_time = last_seconds + duration
last_seconds += duration
labels.append(AudacityLabel(segment_start_time, segment_end_time, all_files_dict[v]))
files_to_concatenate.append(v)
audacity_labels = Util.combine_adjacent_labels_of_the_same_class(labels)
Util.write_audacity_labels(audacity_labels, "gtzan_combined.txt")
command = []
command.append("sox")
command.extend(files_to_concatenate)
command.append("gtzan_combined.wav")
subprocess.check_output(command)
subprocess.call(['chmod', '-R', '777', './music_speech'])
shutil.rmtree("./music_speech")
def x(youtube_video_id):
job = Job.query.filter_by(video_id=youtube_video_id).first()
if job is not None:
try:
job.start_time = datetime.utcnow()
subprocess.check_call([
'youtube-dl', '-f', '18', '--write-thumbnail', '-o',
'videos/%(id)s.%(ext)s',
'https://www.youtube.com/watch?v=%s' % youtube_video_id
])
copyfile('videos/%s.jpg' % youtube_video_id,
'static/img/thumbs/%s.jpg' % youtube_video_id)
set_state(State.VIDEO_DOWNLOADED, db, job)
subprocess.check_call([
'ffmpeg', '-y', '-i',
'videos/%s.mp4' % youtube_video_id, '-ar', '16000', '-ac', '1',
'audios/%s.wav' % youtube_video_id
])
set_state(State.AUDIO_EXTRACTED, db, job)
y, sr = librosa.load("audios/%s.wav" % youtube_video_id, sr=16000)
D = librosa.amplitude_to_db(librosa.stft(y), ref=np.max)
D = np.flipud(D)
D8 = (((D - D.min()) / (D.max() - D.min())) * 255.9).astype(
np.uint8)
img = Image.fromarray(D8)
img = img.resize((D.shape[1], 64))
img.save("static/img/waveforms/%s.jpg" % youtube_video_id)
duration = librosa.get_duration(y=y, sr=sr)
job.waveform_width = D.shape[1]
job.duration = duration
set_state(State.WAVEFORM_GENERATED, db, job)
if duration > 600:
raise Exception("Video duration greated than 10 min (600 sec)")
# audio based segmentation
generate_audio_based_segmentation(
os.path.abspath('audios/%s.wav' % youtube_video_id),
15,
20,
256,
128,
0.2,
os.path.abspath('models/weights.h5'),
os.path.abspath('models/scaler.pickle'),
1024,
3,
1024,
youtube_video_id,
os.path.abspath('static/lbls/audio'),
clusters=job.number_of_speakers)
set_state(State.AUDIO_DATA_ANALYSED, db, job)
steps.mfcc.generate_audio_based_segmentation(
os.path.abspath('audios/%s.wav' % youtube_video_id),
15,
20,
256,
128,
0.2,
os.path.abspath('models/weights.h5'),
os.path.abspath('models/scaler.pickle'),
1024,
3,
1024,
youtube_video_id,
os.path.abspath('static/lbls/mfcc'),
clusters=job.number_of_speakers)
set_state(State.MFCC_ANALYSED, db, job)
# face based segmentation
extract_images_from_video(
os.path.abspath('videos/%s.mp4' % youtube_video_id),
os.path.abspath('video_frames'))
generate_face_based_segmentation(
youtube_video_id,
os.path.abspath('video_frames/%s' % youtube_video_id),
os.path.abspath('static/lbls/image'), job.number_of_speakers,
os.path.abspath(
'models/shape_predictor_68_face_landmarks.dat'),
os.path.abspath(
'models/dlib_face_recognition_resnet_model_v1.dat'))
set_state(State.IMAGE_DATA_ANALYSED, db, job)
# fusion
mapping_face_to_voice = calculate_fusion(
youtube_video_id, os.path.abspath('static/lbls/fusion'),
Util.read_audacity_labels(
os.path.abspath('static/lbls/audio/%s.txt' %
youtube_video_id)),
Util.read_audacity_labels(
os.path.abspath('static/lbls/image/%s.txt' %
youtube_video_id)), duration)
job.mapping_face_to_voice = mapping_face_to_voice
set_state(State.FUSION_APPLIED, db, job)
job.end_time = datetime.utcnow()
set_state(State.DONE, db, job)
except Exception as e:
print(e)
job.end_time = datetime.utcnow()
set_state(State.ERROR, db, job, str(e))
raise e
def generate_face_based_segmentation(youtube_video_id, images_dir, lbls_dir,
faces, predictor_path,
face_rec_model_path, tmp_dir):
images_raw = glob(os.path.join(images_dir, "*.jpg"))
images_raw.sort()
# images_raw = images_raw[0:100]
images = [
images_raw[i] for i in range(0, len(images_raw), FRAMES_PER_STEP)
]
print(images)
timestamps = [i * (FRAMES_PER_STEP / 25.0) for i in range(0, len(images))]
print(timestamps)
detector = get_frontal_face_detector()
sp = dlib.shape_predictor(predictor_path)
facerec = face_recognition_model_v1(face_rec_model_path)
embeddings = []
embeddings_timestamps = []
landmarks_parts = []
landmarks_rect = []
embeddings_pickle = os.path.join(tmp_dir, "embeddings.npy")
embeddings_timestamps_pickle = os.path.join(tmp_dir,
"embeddings_timestamps.npy")
if not os.path.isfile(embeddings_pickle) or not os.path.isfile(
embeddings_timestamps_pickle):
for frame_no, f in enumerate(images):
print("Processing file: {}".format(f))
img = io.imread(f)
dets = detector(img, 1)
print("Number of faces detected: {}".format(len(dets)))
for k, d in enumerate(dets):
shape = sp(img, d)
face_descriptor = facerec.compute_face_descriptor(img, shape)
embeddings.append(face_descriptor)
embeddings_timestamps.append(timestamps[frame_no])
landmarks_parts.append(shape.parts())
landmarks_rect.append(shape.rect)
embeddings = numpy.array(embeddings)
embeddings_timestamps = numpy.array(embeddings_timestamps)
numpy.save(embeddings_pickle, embeddings)
numpy.save(embeddings_timestamps_pickle, embeddings_timestamps)
else:
embeddings = numpy.load(embeddings_pickle)
embeddings_timestamps = numpy.load(embeddings_timestamps_pickle)
print(embeddings.shape)
print(embeddings_timestamps.shape)
if len(embeddings) == 0:
Util.write_audacity_labels([],
os.path.join(
lbls_dir,
youtube_video_id + ".image.txt"))
return
kmeans = KMeans(n_clusters=faces)
kmeans.fit(embeddings)
predictions = numpy.array(kmeans.labels_.tolist())
df = pd.DataFrame({
"timestamps": embeddings_timestamps.tolist(),
"predictions": predictions
})
timestamps = []
classes = []
for key, group in df.groupby(['timestamps']):
timestamps.append(key)
classes.append(",".join(
[str(i) for i in sorted(group['predictions'].tolist())]))
lbls = Util.generate_labels_from_classifications(classes, timestamps)
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".image.txt"))
import itertools
import pickle
import numpy as np
from sac.util import Util
import os
import feat
from sklearn.svm import SVC
DATASETS = os.path.abspath("../datasets")
features = Util.read_feature_names_from_file(
os.path.join(DATASETS, "featureplans/featureplan"))
features1 = ["zcr", "flux", "spectral_rollof", "energy_stats"]
features2 = ["mfcc_stats"]
features3 = ["spectral_flatness_per_band"]
features4 = features1 + features2 + features3
data = feat.get_features(features4, DATASETS, pca=False)
TRAIN = ["mirex", "labrosa", "gtzan"]
model = SVC()
X = np.vstack((data["x_" + i] for i in TRAIN))
Y = list(itertools.chain.from_iterable([data["y_" + i] for i in TRAIN]))
model.fit(X, Y)
with open("pickled/model.pickle", 'w') as f:
pickle.dump(model, f)
import itertools
import pickle
import numpy as np
from sac.util import Util
import os
import feat
from sklearn.svm import SVC
DATASETS = os.path.abspath("../datasets")
features = Util.read_feature_names_from_file(os.path.join(DATASETS, "featureplans/featureplan"))
features1 = ["zcr", "flux", "spectral_rollof", "energy_stats"]
features2 = ["mfcc_stats"]
features3 = ["spectral_flatness_per_band"]
features4 = features1 + features2 + features3
data = feat.get_features(features4, DATASETS, pca=False)
TRAIN = ["mirex", "labrosa", "gtzan"]
model = SVC()
X = np.vstack((data["x_" + i] for i in TRAIN))
Y = list(itertools.chain.from_iterable([data["y_" + i] for i in TRAIN]))
model.fit(X, Y)
with open("pickled/model.pickle", 'w') as f:
pickle.dump(model, f)
def generate_face_based_segmentation(youtube_video_id, images_dir, lbls_dir,
faces, predictor_path,
face_rec_model_path):
images_raw = glob(os.path.join(images_dir, "*.jpg"))
images_raw.sort()
# images_raw = images_raw[0:100]
images = [
images_raw[i] for i in range(0, len(images_raw), FRAMES_PER_STEP)
]
print(images)
timestamps = [i * (FRAMES_PER_STEP / 25.0) for i in range(0, len(images))]
print(timestamps)
detector = dlib.get_frontal_face_detector()
sp = dlib.shape_predictor(predictor_path)
facerec = dlib.face_recognition_model_v1(face_rec_model_path)
embeddings = []
embeddings_timestamps = []
landmarks_parts = []
landmarks_rect = []
for frame_no, f in enumerate(images):
print("Processing file: {}".format(f))
img = io.imread(f)
dets = detector(img, 1)
print("Number of faces detected: {}".format(len(dets)))
for k, d in enumerate(dets):
shape = sp(img, d)
face_descriptor = facerec.compute_face_descriptor(img, shape)
embeddings.append(face_descriptor)
embeddings_timestamps.append(timestamps[frame_no])
landmarks_parts.append(shape.parts())
landmarks_rect.append(shape.rect)
embeddings = numpy.array(embeddings)
embeddings_timestamps = numpy.array(embeddings_timestamps)
print(embeddings.shape)
print(embeddings_timestamps.shape)
if len(embeddings) == 0:
Util.write_audacity_labels([],
os.path.join(lbls_dir,
youtube_video_id + ".txt"))
return
kmeans = KMeans(n_clusters=faces)
kmeans.fit(embeddings)
predictions = numpy.array(kmeans.labels_.tolist())
df = pd.DataFrame({
"timestamps": embeddings_timestamps.tolist(),
"predictions": predictions
})
timestamps = []
classes = []
for key, group in df.groupby(['timestamps']):
timestamps.append(key)
classes.append(",".join(
[str(i) for i in sorted(group['predictions'].tolist())]))
lbls = Util.generate_labels_from_classifications(classes, timestamps)
json_lbls = []
for lbl in lbls:
json_lbls.append({
"start_seconds": lbl.start_seconds,
"end_seconds": lbl.end_seconds,
"label": lbl.label
})
with open(os.path.join(lbls_dir, youtube_video_id + ".json"),
'w') as outfile:
json.dump(json_lbls, outfile)
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".txt"))
# if __name__ == '__main__':
#
# extract_images_from_video("/Users/nicktgr15/workspace/speaker_diarisation_poc/src/videos/Unamij6z1io.mp4",
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/video_frames")
#
# generate_face_based_segmentation(
# "Unamij6z1io",
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/video_frames/Unamij6z1io",
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/static/lbls/image",
# 4,
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/models/shape_predictor_68_face_landmarks.dat",
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/models/dlib_face_recognition_resnet_model_v1.dat"
# )
def generate_audio_based_segmentation(audio_file,
w,
h,
embedding_size,
lstm_nodes,
dropout,
weights_filename,
scaler_filename,
window_size,
step,
hop_size,
youtube_video_id,
lbls_dir,
clusters=4,
sr=16000):
vad = Vad()
vad_lbls = vad.detect_voice_segments(audio_file)
with open(scaler_filename, 'rb') as f:
scaler = pickle.load(f)
model, intermediate = get_lstm_siamese((w, h), embedding_size, lstm_nodes,
dropout)
model.load_weights(weights_filename)
feature_extractor = FeatExtractorMFCC(window_size,
hop_size,
w,
sr,
h,
step=step)
X, timestamps = feature_extractor.extract(audio_file)
timestamps = numpy.array(timestamps)
window = timestamps[1] - timestamps[0]
frame_predictions = []
for k, timestamp in enumerate(timestamps):
found = False
for lbl in vad_lbls:
if lbl.start_seconds <= timestamp <= lbl.end_seconds - window: # need the window end to fit in the label
frame_predictions.append(lbl.label)
found = True
break
if not found:
frame_predictions.append('non_speech')
frame_predictions = numpy.array(frame_predictions)
print(frame_predictions.shape)
print(timestamps.shape)
speech_indices = numpy.where(frame_predictions == 'speech')
X_speech = X[speech_indices]
X = X_speech.reshape((X_speech.shape[0] * w, h))
X = scaler.transform(X)
X = X.reshape(-1, w, h)
original_embeddings = intermediate.predict(X)
clustering_algorithm = KMeans(n_clusters=clusters)
reducted_embeddings = original_embeddings
predictions = clustering_algorithm.fit_predict(reducted_embeddings)
for k, speech_index in enumerate(speech_indices[0]):
frame_predictions[speech_index] = predictions[k]
lbls = Util.generate_labels_from_classifications(frame_predictions,
timestamps)
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".audio.txt"))
all_files_dict[f] = "m"
random.seed(1111)
all_files_random_keys = random.sample(all_files_dict.keys(),
len(all_files_dict.keys()))
last_seconds = 0
files_to_concatenate = []
labels = []
for v in all_files_random_keys:
duration = float(subprocess.check_output(["soxi", "-D", v]).strip())
segment_start_time = last_seconds
segment_end_time = last_seconds + duration
last_seconds += duration
labels.append(
AudacityLabel(segment_start_time, segment_end_time, all_files_dict[v]))
files_to_concatenate.append(v)
audacity_labels = Util.combine_adjacent_labels_of_the_same_class(labels)
Util.write_audacity_labels(audacity_labels, "gtzan_combined.txt")
command = []
command.append("sox")
command.extend(files_to_concatenate)
command.append("gtzan_combined.wav")
subprocess.check_output(command)
subprocess.call(['chmod', '-R', '777', './music_speech'])
shutil.rmtree("./music_speech")
def test_parse_yaafe_header_double_stats(self):
header = Util.parse_yaafe_header(self.double_stats_csv)
self.assertEqual(22050, header['samplerate'])
self.assertEqual(460800, header['effective_step_size'])
def generate_audio_based_segmentation(audio_file,
w,
h,
embedding_size,
lstm_nodes,
dropout,
weights_filename,
scaler_filename,
window_size,
step,
hop_size,
youtube_video_id,
lbls_dir,
clusters=4,
sr=16000):
vad = Vad()
vad_lbls = vad.detect_voice_segments(audio_file)
# model.load_weights(weights_filename)
# feature_extractor = FeatExtractorMFCC(window_size, hop_size, w, sr, h, step=step)
# X, timestamps = feature_extractor.extract(audio_file)
# timestamps = numpy.array(timestamps)
y, sr = librosa.load(audio_file, sr=sr)
X = mfcc(y, sr=sr, n_mfcc=h, n_fft=window_size, hop_length=hop_size)
timestamps = [k * hop_size / sr for k in range(0, X.shape[1])]
timestamps = numpy.array(timestamps)
window = timestamps[1] - timestamps[0]
frame_predictions = []
for k, timestamp in enumerate(timestamps):
found = False
for lbl in vad_lbls:
if lbl.start_seconds <= timestamp <= lbl.end_seconds - window: # need the window end to fit in the label
frame_predictions.append(lbl.label)
found = True
break
if not found:
frame_predictions.append('non_speech')
frame_predictions = numpy.array(frame_predictions)
speech_indices = numpy.where(frame_predictions == 'speech')
X_speech = X[:, speech_indices]
X_speech = X_speech.reshape((X_speech.shape[0], X_speech.shape[2]))
print(X_speech.shape)
X_speech = X_speech.transpose()
print(X_speech.shape)
timestamps_speech = timestamps[speech_indices]
# Util.write_audacity_labels(Util.generate_labels_from_classifications(frame_predictions, timestamps),
# "vad_preds_quant.txt")
# X = X_speech.reshape((X_speech.shape[0] * w, h))
# X = scaler.transform(X)
# X = X.reshape(-1, w, h)
# original_embeddings = intermediate.predict(X)
clustering_algorithm = GaussianMixture(n_components=clusters,
max_iter=1000,
n_init=3)
# if visualise:
#
# embeddings, y, classes, new_timestamps = Util.get_annotated_data_x_y(timestamps_speech, original_embeddings,
# lbls_fixed)
# le = preprocessing.LabelEncoder()
# y = le.fit_transform(y)
#
# tsne = TSNE()
# two_dimensional = tsne.fit_transform(embeddings)
#
# # pca = PCA(n_components=2)
# # two_dimensional = pca.fit_transform(embeddings)
#
# # pca2 = PCA(n_components=20)
# # pca_embeddings = pca2.fit_transform(embeddings)
#
# clustering_algorithm.fit(two_dimensional)
# predictions = clustering_algorithm.predict(two_dimensional)
#
# # kmeans = KMeans(n_clusters=CLUSTERS)
# # kmeans.fit(embeddings)
#
# plt.figure(figsize=(10, 6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=y, marker='.')
#
# plt.figure(figsize=(10, 6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=predictions, marker='.')
#
# else:
tsne = TSNE(n_components=2, init='pca')
two_dimensional = tsne.fit_transform(X_speech)
# original_embeddings = scale((original_embeddings))
# pca = PCA(n_components=2)
# two_dimensional = pca.fit_transform(original_embeddings)
# pca2 = PCA(n_components=3)
# pca_embeddings = pca2.fit_transform(original_embeddings)
clustering_algorithm.fit(two_dimensional)
predictions = clustering_algorithm.predict(two_dimensional)
# kmeans = KMeans(n_clusters=CLUSTERS)
# kmeans.fit(two_dimensional)
# plt.figure(figsize=(10,10))
# plt.imshow(calculate_similarity_matrix(two_dimensional, metric='euclidean'), cmap='gray')
# plt.figure(figsize=(10,6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=predictions, marker='.')
# plt.figure(figsize=(10,6))
# plt.scatter(two_dimensional[:, 0], pca_embeddings[:, 1], c=kmeans.labels_.tolist(), marker='.')
# predictions = kmeans.labels_.tolist()
for k, speech_index in enumerate(speech_indices[0]):
frame_predictions[speech_index] = predictions[k]
lbls = Util.generate_labels_from_classifications(frame_predictions,
timestamps)
json_lbls = []
for lbl in lbls:
json_lbls.append({
"start_seconds": lbl.start_seconds,
"end_seconds": lbl.end_seconds,
"label": lbl.label
})
with open(os.path.join(lbls_dir, youtube_video_id + ".json"),
'w') as outfile:
json.dump(json_lbls, outfile)
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".txt"))
# if __name__ == '__main__':
# generate_audio_based_segmentation(
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/audios/Unamij6z1io.wav",
# 15, 20, 256, 128, 0.2,
# os.path.abspath('models/weights.h5'),
# os.path.abspath('models/scaler.pickle'),
# 1024, 3, 1024, "xxx",
# "/Users/nicktgr15/workspace/speaker_diarisation_poc/src/static/lbls/mfcc",
# clusters=4
# )
def generate_audio_based_segmentation(audio_file,
w,
h,
embedding_size,
lstm_nodes,
dropout,
weights_filename,
scaler_filename,
window_size,
step,
hop_size,
youtube_video_id,
lbls_dir,
clusters=4,
sr=16000):
vad = Vad()
vad_lbls = vad.detect_voice_segments(audio_file)
with open(scaler_filename, 'rb') as f:
scaler = pickle.load(f)
model, intermediate = get_lstm_siamese((w, h), embedding_size, lstm_nodes,
dropout)
model.load_weights(weights_filename)
feature_extractor = FeatExtractorMFCC(window_size,
hop_size,
w,
sr,
h,
step=step)
X, timestamps = feature_extractor.extract(audio_file)
timestamps = numpy.array(timestamps)
window = timestamps[1] - timestamps[0]
frame_predictions = []
for k, timestamp in enumerate(timestamps):
found = False
for lbl in vad_lbls:
if lbl.start_seconds <= timestamp <= lbl.end_seconds - window: # need the window end to fit in the label
frame_predictions.append(lbl.label)
found = True
break
if not found:
frame_predictions.append('non_speech')
frame_predictions = numpy.array(frame_predictions)
print(frame_predictions.shape)
print(timestamps.shape)
speech_indices = numpy.where(frame_predictions == 'speech')
X_speech = X[speech_indices]
timestamps_speech = timestamps[speech_indices]
# Util.write_audacity_labels(Util.generate_labels_from_classifications(frame_predictions, timestamps),
# "vad_preds_quant.txt")
X = X_speech.reshape((X_speech.shape[0] * w, h))
X = scaler.transform(X)
X = X.reshape(-1, w, h)
original_embeddings = intermediate.predict(X)
clustering_algorithm = GaussianMixture(n_components=clusters,
max_iter=1000,
n_init=3)
# if visualise:
#
# embeddings, y, classes, new_timestamps = Util.get_annotated_data_x_y(timestamps_speech, original_embeddings,
# lbls_fixed)
# le = preprocessing.LabelEncoder()
# y = le.fit_transform(y)
#
# tsne = TSNE()
# two_dimensional = tsne.fit_transform(embeddings)
#
# # pca = PCA(n_components=2)
# # two_dimensional = pca.fit_transform(embeddings)
#
# # pca2 = PCA(n_components=20)
# # pca_embeddings = pca2.fit_transform(embeddings)
#
# clustering_algorithm.fit(two_dimensional)
# predictions = clustering_algorithm.predict(two_dimensional)
#
# # kmeans = KMeans(n_clusters=CLUSTERS)
# # kmeans.fit(embeddings)
#
# plt.figure(figsize=(10, 6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=y, marker='.')
#
# plt.figure(figsize=(10, 6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=predictions, marker='.')
#
# else:
tsne = TSNE(n_components=2, init='pca')
two_dimensional = tsne.fit_transform(original_embeddings)
# original_embeddings = scale((original_embeddings))
# pca = PCA(n_components=2)
# two_dimensional = pca.fit_transform(original_embeddings)
# pca2 = PCA(n_components=3)
# pca_embeddings = pca2.fit_transform(original_embeddings)
clustering_algorithm.fit(two_dimensional)
predictions = clustering_algorithm.predict(two_dimensional)
# kmeans = KMeans(n_clusters=CLUSTERS)
# kmeans.fit(two_dimensional)
# plt.figure(figsize=(10,10))
# plt.imshow(calculate_similarity_matrix(two_dimensional, metric='euclidean'), cmap='gray')
# plt.figure(figsize=(10,6))
# plt.scatter(two_dimensional[:, 0], two_dimensional[:, 1], c=predictions, marker='.')
# plt.figure(figsize=(10,6))
# plt.scatter(two_dimensional[:, 0], pca_embeddings[:, 1], c=kmeans.labels_.tolist(), marker='.')
# predictions = kmeans.labels_.tolist()
for k, speech_index in enumerate(speech_indices[0]):
frame_predictions[speech_index] = predictions[k]
lbls = Util.generate_labels_from_classifications(frame_predictions,
timestamps)
json_lbls = []
for lbl in lbls:
json_lbls.append({
"start_seconds": lbl.start_seconds,
"end_seconds": lbl.end_seconds,
"label": lbl.label
})
with open(os.path.join(lbls_dir, youtube_video_id + ".json"),
'w') as outfile:
json.dump(json_lbls, outfile)
Util.write_audacity_labels(
lbls, os.path.join(lbls_dir, youtube_video_id + ".txt"))
