def main():
# man 900 middle 750 default 500 frame_length = 900
parser = argparse.ArgumentParser()
parser.add_argument('-s', '--speed', type=float, default=1.)
parser.add_argument('-t', '--time', type=float, default=-10.)
parser.add_argument('-o', '--output', type=str, default='output.wav')
parser.add_argument('-i', '--input', type=str, default='input.wav')
args = parser.parse_args()
parameters = {}
input_filename = args.input
output_filename = args.output
if not os.path.isfile(input_filename):
raise RuntimeError('no input file')
x, fs = core.load(input_filename)
#f0, sp, ap = pw.wav2world(x, fs)
frame_length = 1500 #100000 // int(calculateF0(f0)) // 2 * 2
y, sr = core.load(input_filename, sr=fs)
onset_frames = onset.onset_detect(x,
sr=sr,
wait=1,
pre_avg=1,
post_avg=1,
pre_max=1,
post_max=1)
onset_times = librosa.frames_to_time(onset_frames)
plt.plot(y)
for i in onset_times:
plt.plot([i * 22050, i * 22050], [-1, 1], color="red")
S = librosa.stft(x)
logS = librosa.amplitude_to_db(abs(S))
plt.savefig('woman.png')
if args.time < 0:
parameters['origin_time'] = core.get_duration(y, sr)
parameters['convert_time'] = parameters['origin_time'] / args.speed
else:
parameters['origin_time'] = core.get_duration(y, sr)
parameters['convert_time'] = args.time
parameters['sample_rate'] = sr
parameters['frame_length'] = int(fs / 22050 * frame_length)
#if parameters['convert_time'] / parameters['origin_time'] > 0.8:
convert_upper_threshold(input_filename, output_filename, parameters)
def upload_tracks(request, project_id):
message = ""
project = Project.objects.get(id=project_id)
if request.method == "POST":
form = MultipleFileFieldForm(request.POST, request.FILES)
current_files = project.audiotrack_set.values_list("name", flat=True)
if form.is_valid():
files = request.FILES.getlist("file_field")
non_audio = 0
duplicate = 0
success = 0
for f in files:
if not f.name.endswith((".mp3", ".ogg", ".wav", ".flac")):
non_audio += 1
continue
if f.name in current_files:
duplicate += 1
continue
local_file = os.path.join(settings.MEDIA_ROOT, f.name)
with open(local_file, 'wb+') as destination:
for chunk in f.chunks():
destination.write(chunk)
try:
duration = get_duration(filename=local_file)
except Exception:
duration = 0
AudioTrack.objects.create(
name=f.name,
file=os.path.join(settings.MEDIA_URL, f.name),
format=os.path.splitext(f.name)[1][1:],
project=project,
duration=duration)
success += 1
message = f"{success} files successfully uploaded"
if non_audio:
message += f", {non_audio} non-audio files rejected"
if duplicate:
message += f", {duplicate} duplicate files rejected"
form = MultipleFileFieldForm()
return render(request, 'annotate/upload.html', {
'form': form,
'project': project,
'message': message
})
def get_next_clip(all_files):
# select only clips that are available
all_files = [a for a in all_files if a["elapsed"] is False]
# if no more files are available, exit
if len(all_files) == 0:
return None
# randomly pick a file
selected_idx = np.random.randint(0, len(all_files))
selected_file = all_files[selected_idx]
file_path = selected_file["file_path"]
# load only part of the clip
clip, sr = librosa.load(file_path,
sr=args.sample_rate,
offset=selected_file["time_elapsed"].seconds,
duration=args.max_clip_len)
assert sr == args.sample_rate
clip_len = timedelta(seconds=get_duration(clip, sr=sr))
if clip_len.seconds == 0:
selected_file["elapsed"] = True
return timedelta(seconds=0, hours=0, minutes=0)
elapsed_time = selected_file["time_elapsed"].seconds
# print("Clip ID: ", elapsed_time, "Clip Len:", clip_len,
# "Time Elapsed: ", selected_file["time_elapsed"])
selected_file["time_elapsed"] += clip_len
split_clip_id = "{}_{}_{}".format(selected_file["book_id"],
selected_file["clip_id"], elapsed_time)
dest_path = os.path.join(output_dir, "{}.wav".format(split_clip_id))
log.debug("Saving clip to {}".format(dest_path))
librosa.output.write_wav(dest_path, clip, args.sample_rate)
return clip_len
def extract_feature(self, song_fname):
song, _ = self.load_song(song_fname)
duration = int(get_duration(song))
n = song.shape[0]
num_windows = int(np.floor((duration - self.wsize) / self.stride)) + 20
feats = np.zeros((num_windows, self.feat_dim), dtype=np.float32)
start = 0
counter = 0
while True:
end = start + int(self.wsize * self.sr)
if end >= n:
break
window = song[start:end]
feat_ = self._calc_all_feat(window)
feats[counter] = feat_
counter += 1
start += int(self.stride * self.sr)
if start >= n:
break
if counter < num_windows:
feats = feats[:counter, :]
return feats
start = datetime.now()
# Get speaker ID
split, _, speaker, _ = fpath.split(os.sep)[-4:]
if speaker in all_speakers:
speaker_id = all_speakers.index(speaker)
else:
speaker_id = len(all_speakers)
all_speakers.append(speaker)
# Get duration
if debug:
print('\n==== DEBUG ====')
print(fpath)
try:
duration = str(timedelta(seconds=get_duration(filename=fpath)))
if debug:
print('File duration: {}'.format(duration))
except:
print('UNABLE TO GET DURATION')
raise
# Chunk into segments with speech audio
times, segs = VAD_chunk(2, fpath)
if segs == []:
print('No voice activity detected')
continue
if debug:
print('{} - {:,} segments'.format(datetime.now() - start,
len(segs)))
transforms.ToTensor(), normalize ]) categories = utils.categories ### ImageNet Categories places_categories = placesCNN_basic.classes ### Places Categories fps = 24 nb_frames = 1 nbsec = 1.49 n_obj = 3 beg_film = 0 end_film = np.floor(get_duration(filename=wavfile)) allpreds = [] onsets = [] model_imagenet = resnet18(pretrained=True) model_imagenet.eval() model_places = placesCNN_basic.model.eval() ### Define and register hook for extracting output feature map places_fm = [] places_proba = [] def get_fm_places(m, i, o):
#!/usr/bin/env python
from sklearn.metrics.pairwise import pairwise_distances
import scipy
from librosa.core import get_duration
import numpy as np
import h5py
from scipy.fftpack import dct, idct
from utils import (WINDOW_SIZE, HOP_SIZE, SAMPLE_RATE, COEFS, TIMBRE_GROUP)
ONE_FRAME = get_duration(sr=SAMPLE_RATE,
n_fft=WINDOW_SIZE,
hop_length=HOP_SIZE,
S=np.zeros((1, 2)))
def beat_similarity(vecs):
return 1.0 - pairwise_distances(vecs, metric='cosine')
def normed_diags(B, max_lag):
B_diags = np.asarray([B.trace(i) for i in range(max_lag)])
B_diags -= np.min(B_diags)
return B_diags / np.max(B_diags)
def fft_autocorrelate(S):
B_fft = scipy.signal.fftconvolve(S, S[::-1, ::-1], mode='full')
B_fft = B_fft[S.shape[0]:, S.shape[1]:]
return B_fft
with open(os.path.join(meta_path, cl) + ".json") as reader:
meta_data = json.load(reader)
meta_data = {d["book_id"]: d for d in meta_data}
book_ids = defaultdict(
lambda: timedelta(hours=0, minutes=0, seconds=0))
author_ids = defaultdict(
lambda: timedelta(hours=0, minutes=0, seconds=0))
for idx, file_id in enumerate(all_files):
file_path = os.path.join(cl_path, file_id)
sound, sample_rate = librosa.load(file_path, sr=None)
assert sample_rate == assigned_sample_rate
current_time += timedelta(
seconds=get_duration(sound, sr=sample_rate))
book_id = file_id.split("_")[0]
book_ids[book_id] += timedelta(
seconds=get_duration(sound, sr=sample_rate))
author_ids[meta_data[book_id]["reader_url"]] += timedelta(
seconds=get_duration(sound, sr=sample_rate))
if idx % 100 == 0:
print("\t{} of {}".format(idx, len(all_files)))
print("{} Books".format(cl))
for book_id in book_ids:
print("\t{}: Time: {}".format(book_id, book_ids[book_id]))
book_times.append({
"language": cl,
# collect all wav files
cwd = os.getcwd()
stations = filter(os.path.isdir, os.listdir(cwd))
stations = [f for f in stations if not f.startswith('.')]
audio_paths = []
for station in stations:
data_path = f'{cwd}/{station}'
files = os.listdir(data_path)
wav_names = [f for f in files if f.endswith('.wav')]
wav_paths = [f'{data_path}/{f}' for f in wav_names]
audio_paths.extend(wav_paths)
print(f'we found {len(audio_paths)} audio files in total')
lens = [get_duration(filename=f) for f in audio_paths]
print(f'their total play time amounts to {sum(lens) / 60 / 60:.2f} hours')
# collect all aup files
labeled_data = []
for station in stations:
data_path = f'{cwd}/{station}'
files = os.listdir(data_path)
audacity_projects = [f for f in files if f.endswith('.aup')]
project_paths = [f'{data_path}/{aup}' for aup in audacity_projects]
station_data = [extract_aup(i, data_path, station, verbose=0) for i in project_paths]
labeled_data.extend(station_data)
print(f'{len(labeled_data)} instances have labels provided')
# infer amount of detections