def test_import_tracks(self, corpus):
importing_tracks = [
tracks.FileTrack('a', '/some/path.wav'),
tracks.FileTrack('b', '/some/other/path.wav'),
tracks.FileTrack('existing_file', '/some/otherer/path.wav'),
]
idx_mapping = corpus.import_tracks(importing_tracks)
assert corpus.num_tracks == 4
assert 'a' in corpus.tracks.keys()
assert corpus.tracks['a'].path == '/some/path.wav'
assert 'b' in corpus.tracks.keys()
assert corpus.tracks['b'].path == '/some/other/path.wav'
assert 'existing_file_1' in corpus.tracks.keys()
assert corpus.tracks[
'existing_file_1'].path == '/some/otherer/path.wav'
assert len(idx_mapping) == 3
assert 'a' in idx_mapping['a'].idx
assert 'b' in idx_mapping['b'].idx
assert idx_mapping['existing_file'].idx == 'existing_file_1'
def test_exports_wavs_from_container_tracks(self, writer, tmpdir):
path = tmpdir.strpath
container_ds_path = os.path.join(path, 'container_ds')
out_path = os.path.join(path, 'export')
ds = resources.create_dataset()
ds.relocate_audio_to_single_container(container_ds_path)
writer.save(ds, out_path)
print(os.listdir(out_path))
track_path = os.path.join(out_path, 'audio', 'wav-1.wav')
track = tracks.FileTrack(None, track_path)
assert os.path.isfile(track_path)
assert track.duration == pytest.approx(2.5951875)
assert np.allclose(track.read_samples(),
ds.tracks['wav-1'].read_samples(),
atol=1e-05)
track_path = os.path.join(out_path, 'audio', 'wav_2.wav')
track = tracks.FileTrack(None, track_path)
assert os.path.isfile(track_path)
assert track.duration == pytest.approx(2.5951875)
track_path = os.path.join(out_path, 'audio', 'wav_3.wav')
track = tracks.FileTrack(None, track_path)
assert os.path.isfile(track_path)
assert track.duration == pytest.approx(2.5951875)
track_path = os.path.join(out_path, 'audio', 'wav_4.wav')
track = tracks.FileTrack(None, track_path)
assert os.path.isfile(track_path)
assert track.duration == pytest.approx(2.5951875)
def test_encode_label_ends_at_utterance_end(self):
track = tracks.FileTrack('file1',
resources.sample_wav_file('med_len.wav'))
utt = tracks.Utterance('utt1', track, start=3, end=14)
ll = annotations.LabelList(labels=[
annotations.Label('speech', 0, 4),
annotations.Label('music', 4, 9),
annotations.Label('speech', 9, float('inf')),
])
utt.set_label_list(ll)
enc = encoding.FrameHotEncoder(['music', 'speech', 'noise'],
'default',
frame_settings=units.FrameSettings(
32000, 16000),
sr=16000)
actual = enc.encode_utterance(utt)
expected = np.array([
[0, 1, 0],
[0, 1, 0],
[0, 1, 0],
[1, 1, 0],
[1, 0, 0],
[1, 0, 0],
[1, 0, 0],
[1, 0, 0],
[1, 1, 0],
[0, 1, 0],
]).astype(np.float32)
assert np.array_equal(expected, actual)
def relocate_audio_to_wav_files(self, target_path):
"""
Copies every track to its own wav file in the given folder.
Every track will be stored at ``target_path/track_id.wav``.
"""
if not os.path.isdir(target_path):
os.makedirs(target_path)
new_tracks = {}
# First create a new container track for all existing tracks
for track in self.tracks.values():
track_path = os.path.join(target_path, '{}.wav'.format(track.idx))
sr = track.sampling_rate
samples = track.read_samples()
audio.write_wav(track_path, samples, sr=sr)
new_track = tracks.FileTrack(track.idx, track_path)
new_tracks[track.idx] = new_track
# Update track list of corpus
self._tracks = new_tracks
# Update utterances to point to new tracks
for utterance in self.utterances.values():
new_track = self.tracks[utterance.track.idx]
utterance.track = new_track
def test_process_track_online(self, processor, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(174)
librosa.output.write_wav(wav_path, wav_content, 16000)
track = tracks.FileTrack('idx', wav_path)
chunks = list(
processor.process_track_online(track,
frame_size=20,
hop_size=10,
chunk_size=8))
assert len(chunks) == 3
assert np.allclose(chunks[0][0], wav_content[0:20], atol=0.0001)
assert np.allclose(chunks[2][-1],
np.pad(wav_content[160:], (0, 6), mode='constant'),
atol=0.0001)
assert chunks[0].dtype == np.float32
assert processor.called_with_sr == [16000, 16000, 16000]
assert processor.called_with_offset == [0, 8, 16]
assert processor.called_with_last == [False, False, True]
assert processor.called_with_utterance == [None, None, None]
assert processor.called_with_corpus == [None, None, None]
def test_does_utt_match_target_format_with_invalid_format_returns_false(self):
file_path = resources.get_resource_path(('audio_formats', 'mp3_2_44_1k_16b.mp3'))
track = tracks.FileTrack('t', file_path)
utt = tracks.Utterance('u', track)
c = conversion.WavAudioFileConverter()
assert not c._does_utt_match_target_format(utt)
def test_does_utt_match_target_format_returns_true(self):
file_path = resources.sample_wav_file('wav_1.wav')
track = tracks.FileTrack('t', file_path)
utt = tracks.Utterance('u', track)
c = conversion.WavAudioFileConverter()
assert c._does_utt_match_target_format(utt)
def test_read_samples(self):
path = resources.sample_wav_file('wav_1.wav')
track = tracks.FileTrack('wav', path)
issuer = issuers.Issuer('toni')
utt = tracks.Utterance('t', track, issuer=issuer, start=1.0, end=2.30)
l1 = annotations.Label('a', 0.15, 0.448)
l2 = annotations.Label('a', 0.5, 0.73)
ll = annotations.LabelList(labels=[l1, l2])
utt.set_label_list(ll)
expected, __ = librosa.core.load(path,
sr=None,
offset=1.15,
duration=0.298)
assert np.array_equal(l1.read_samples(), expected)
expected, __ = librosa.core.load(path,
sr=None,
offset=1.5,
duration=1.73 - 1.5)
print(expected.shape)
print(l2.read_samples().shape)
assert np.array_equal(l2.read_samples(), expected)
def test_compute_online(self):
test_file_path = resources.sample_wav_file('wav_1.wav')
y, sr = librosa.load(test_file_path, sr=None)
# EXPECTED
y_pad = np.pad(y, (0, 1024), mode='constant', constant_values=0)
S = np.abs(
librosa.stft(y_pad, center=False, n_fft=2048, hop_length=1024))**2
S = librosa.feature.melspectrogram(S=S, n_mels=128, sr=sr)
S = librosa.power_to_db(S)
exp_onsets = librosa.onset.onset_strength(S=S, center=False).T
exp_onsets = exp_onsets.reshape(exp_onsets.shape[0], 1)
# ACTUAL
test_file = tracks.FileTrack('idx', test_file_path)
onset = pipeline.OnsetStrength()
onset_gen = onset.process_track_online(test_file,
2048,
1024,
chunk_size=5)
chunks = list(onset_gen)
onsets = np.vstack(chunks)
print(onsets.shape, exp_onsets.shape)
assert np.allclose(onsets, exp_onsets)
def test_split_sets_track(self):
file = tracks.FileTrack('file-1', '/some/path')
utt = tracks.Utterance('utt-1', file, start=0.0, end=10.0)
res = utt.split([5.2])
assert len(res) == 2
assert res[0].track == file
assert res[1].track == file
def test_read_samples_fix_sampling_rate(self, name, audio_path):
audio_path = os.path.join(audio_path, name)
file_obj = tracks.FileTrack('some_idx', audio_path)
expected, __ = librosa.core.load(audio_path, sr=16000, mono=True)
actual = file_obj.read_samples(sr=16000)
assert np.array_equal(actual, expected)
def test_import_utterance_no_track(self, corpus):
importing_utterances = [
tracks.Utterance('a', tracks.FileTrack('notexist', 'notexist'),
corpus.issuers['existing_issuer'], 0, 10)
]
with pytest.raises(ValueError):
corpus.import_utterances(importing_utterances)
def test_process_empty_track_raises_error(self, processor, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(0)
librosa.output.write_wav(wav_path, wav_content, 16000)
file_track = tracks.FileTrack('idx', wav_path)
with pytest.raises(ValueError):
processor.process_track(file_track, frame_size=4096, hop_size=2048, sr=16000)
def test_utt_read_samples(benchmark):
utts = []
wav_path = resources.get_test_resource_path(('wav_files', 'med_len.wav'))
track = tracks.FileTrack('idx', wav_path)
utts.append(tracks.Utterance('uidx', track))
utts.append(tracks.Utterance('uidx', track, start=2.8))
utts.append(tracks.Utterance('uidx', track, end=10.2))
utts.append(tracks.Utterance('uidx', track, start=2.4, end=9.8))
mp3_path = resources.get_test_resource_path(
('audio_formats', 'mp3_2_44_1k_16b.mp3'))
track = tracks.FileTrack('idx', mp3_path)
utts.append(tracks.Utterance('uidx', track))
utts.append(tracks.Utterance('uidx', track, start=2.8))
utts.append(tracks.Utterance('uidx', track, end=4.9))
utts.append(tracks.Utterance('uidx', track, start=0.4, end=4.8))
benchmark(run, utts)
def _load(self, path):
corpus = audiomate.Corpus()
article_paths = sorted(self.get_articles(path))
reader_map = {}
file_map = {}
for article_path in article_paths:
audio_files = self.get_audio_file_info(article_path)
reader_name, reader_gender = self.get_reader_info(article_path)
segments = self.get_segments(article_path)
if reader_name not in reader_map.keys():
speaker = issuers.Speaker('{:0>8}'.format(len(reader_map)),
gender=reader_gender)
reader_map[reader_name] = speaker
corpus.import_issuers(speaker)
else:
speaker = reader_map[reader_name]
for start, end, text in segments:
file_path = self.find_audio_file_for_segment(
start, end, audio_files)
if file_path is not None:
if file_path not in file_map.keys():
track = tracks.FileTrack(
'{:0>10}'.format(len(file_map)), file_path)
file_map[file_path] = track
corpus.import_tracks(track)
else:
track = file_map[file_path]
track_offset = audio_files[file_path]
utt_start = start - track_offset
utt_end = end - track_offset
utt_idx = '{}_{}_{}_{}'.format(speaker.idx, track.idx,
int(start * 1000),
int(end * 1000))
if utt_idx not in self.invalid_utterance_ids:
utt = corpus.new_utterance(utt_idx,
track.idx,
issuer_idx=speaker.idx,
start=utt_start,
end=utt_end)
ll = annotations.LabelList.create_single(
text, audiomate.corpus.LL_WORD_TRANSCRIPT)
utt.set_label_list(ll)
return audiomate.Corpus.from_corpus(corpus)
def test_write_wav(tmpdir):
samples = np.random.random(50000)
sr = 16000
path = os.path.join(tmpdir.strpath, 'audio.wav')
audio.write_wav(path, samples, sr=sr)
assert os.path.isfile(path)
track = tracks.FileTrack('idx', path)
assert np.allclose(samples, track.read_samples(), atol=1.e-04)
def corpus():
corpus = audiomate.Corpus()
ex_file = tracks.FileTrack('existing_file', '../any/path.wav')
ex_issuer = issuers.Issuer('existing_issuer')
ex_utterance = tracks.Utterance('existing_utt', ex_file, issuer=ex_issuer)
corpus.tracks['existing_file'] = ex_file
corpus.issuers['existing_issuer'] = ex_issuer
corpus.utterances['existing_utt'] = ex_utterance
return corpus
def test_read_samples_range(self, name, audio_path):
audio_path = os.path.join(audio_path, name)
file_obj = tracks.FileTrack('some_idx', audio_path)
expected, __ = librosa.core.load(audio_path,
sr=None,
mono=True,
offset=1.0,
duration=1.7)
actual = file_obj.read_samples(offset=1.0, duration=1.7)
assert np.array_equal(actual, expected)
def generate_tracks(n, rand=None):
if rand is None:
rand = random.Random()
items = []
for i in range(n):
track_idx = 'track-{}'.format(i)
path = '/fake/{}.wav'.format(track_idx)
track = tracks.FileTrack(track_idx, path)
items.append(track)
return items
def test_process_track_with_downsampling(self, processor, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(22)
librosa.output.write_wav(wav_path, wav_content, 4)
file_track = tracks.FileTrack('idx', wav_path)
processed = processor.process_track(file_track, frame_size=4, hop_size=2, sr=2)
assert processed.shape == (5, 4)
assert processor.called_with_sr == [2]
assert processor.called_with_offset == [0]
assert processor.called_with_last == [True]
assert processor.called_with_utterance == [None]
assert processor.called_with_corpus == [None]
def test_process_track_smaller_than_frame_size(self, processor, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(22)
librosa.output.write_wav(wav_path, wav_content, 16000)
file_track = tracks.FileTrack('idx', wav_path)
processed = processor.process_track(file_track, frame_size=4096, hop_size=2048, sr=16000)
assert processed.shape == (1, 4096)
assert np.allclose(processed[0], np.pad(wav_content, (0, 4074), mode='constant'), atol=0.0001)
assert processor.called_with_sr == [16000]
assert processor.called_with_offset == [0]
assert processor.called_with_last == [True]
assert processor.called_with_utterance == [None]
assert processor.called_with_corpus == [None]
def setup_method(self):
self.ll_1 = annotations.LabelList(idx='alpha',
labels=[
annotations.Label('a', 3.2, 4.5),
annotations.Label('b', 5.1, 8.9),
annotations.Label(
'c', 7.2, 10.5),
annotations.Label('d', 10.5, 14),
annotations.Label('d', 15, 18)
])
self.ll_2 = annotations.LabelList(idx='bravo',
labels=[
annotations.Label('a', 1.0, 4.2),
annotations.Label('e', 4.2, 7.9),
annotations.Label(
'c', 7.2, 10.5),
annotations.Label('f', 10.5, 14),
annotations.Label('d', 15, 17.3)
])
self.ll_duplicate_idx = annotations.LabelList(
idx='charlie',
labels=[
annotations.Label('t', 1.0, 4.2),
annotations.Label('h', 4.2, 7.9)
])
self.ll_3 = annotations.LabelList(idx='charlie',
labels=[
annotations.Label('a', 1.0, 4.2),
annotations.Label('g', 4.2, 7.9)
])
self.track = tracks.FileTrack('wav',
resources.sample_wav_file('wav_1.wav'))
self.issuer = issuers.Issuer('toni')
self.utt = tracks.Utterance('test',
self.track,
issuer=self.issuer,
start=1.25,
end=1.30,
label_lists=[
self.ll_1, self.ll_2,
self.ll_duplicate_idx, self.ll_3
])
def test_read_frames(self, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(10044)
librosa.output.write_wav(wav_path, wav_content, 16000)
file_obj = tracks.FileTrack('some_idx', wav_path)
data = list(file_obj.read_frames(frame_size=400, hop_size=160))
frames = np.array([x[0] for x in data])
last = [x[1] for x in data]
assert frames.shape == (62, 400)
assert frames.dtype == np.float32
assert np.allclose(frames[0], wav_content[:400], atol=0.0001)
expect = np.pad(wav_content[9760:], (0, 116), mode='constant')
assert np.allclose(frames[61], expect, atol=0.0001)
assert last[:-1] == [False] * (len(data) - 1)
assert last[-1]
def test_process_track(self, processor, tmpdir):
wav_path = os.path.join(tmpdir.strpath, 'file.wav')
wav_content = np.random.random(22)
librosa.output.write_wav(wav_path, wav_content, 4)
file_track = tracks.FileTrack('idx', wav_path)
processed = processor.process_track(file_track, frame_size=4, hop_size=2)
assert processed.shape == (10, 4)
assert processed.dtype == np.float32
assert np.allclose(processed[0], wav_content[0:4], atol=0.0001)
assert np.allclose(processed[9], wav_content[18:22], atol=0.0001)
assert processor.called_with_sr == [4]
assert processor.called_with_offset == [0]
assert processor.called_with_last == [True]
assert processor.called_with_utterance == [None]
assert processor.called_with_corpus == [None]
def test_encode_utterance_takes_lower_index_first(self):
file = tracks.FileTrack('file-idx',
resources.sample_wav_file('wav_1.wav'))
utt = tracks.Utterance('utt-idx', file, start=0, end=5)
ll = annotations.LabelList(labels=[
annotations.Label('music', 0, 3),
annotations.Label('speech', 3, 5)
])
utt.set_label_list(ll)
enc = encoding.FrameOrdinalEncoder(['speech', 'music', 'noise'],
'default',
frame_settings=units.FrameSettings(
32000, 16000),
sr=16000)
actual = enc.encode_utterance(utt)
expected = np.array([1, 1, 0, 0]).astype(np.int)
assert np.array_equal(expected, actual)
def new_file(self, path, track_idx, copy_file=False):
"""
Adds a new audio file to the corpus with the given data.
Parameters:
path (str): Path of the file to add.
track_idx (str): The id to associate the file-track with.
copy_file (bool): If True the file is copied to the data set folder, otherwise the given
path is used directly.
Returns:
FileTrack: The newly added file.
"""
new_file_idx = track_idx
new_file_path = os.path.abspath(path)
# Add index to idx if already existing
if new_file_idx in self._tracks.keys():
new_file_idx = naming.index_name_if_in_list(
new_file_idx, self._tracks.keys())
# Copy file to default file dir
if copy_file:
if not os.path.isdir(self.path):
raise ValueError(
'To copy file the dataset needs to have a path.')
__, ext = os.path.splitext(path)
new_file_folder = os.path.join(self.path, DEFAULT_FILE_SUBDIR)
new_file_path = os.path.join(new_file_folder,
'{}{}'.format(new_file_idx, ext))
os.makedirs(new_file_folder, exist_ok=True)
shutil.copy(path, new_file_path)
# Create file obj
new_file = tracks.FileTrack(new_file_idx, new_file_path)
self._tracks[new_file_idx] = new_file
return new_file
def test_compute_online(self):
# Data: 41523 samples, 16 kHz
# yields 40 frames with frame-size 2048 and hop-size 1024
test_file_path = resources.sample_wav_file('wav_1.wav')
y, sr = librosa.load(test_file_path, sr=None)
# EXPECTED
y_pad = np.pad(y, (0, 1024), mode='constant', constant_values=0)
S = np.abs(librosa.stft(y_pad, center=False, n_fft=2048, hop_length=1024)) ** 2
S = librosa.feature.melspectrogram(S=S, n_mels=128, sr=sr)
S = librosa.power_to_db(S)
onsets = librosa.onset.onset_strength(S=S, center=False)
exp_tgram = librosa.feature.tempogram(onset_envelope=onsets, sr=sr, win_length=4, center=True).T
# ACTUAL
test_file = tracks.FileTrack('idx', test_file_path)
tgram_step = pipeline.Tempogram(win_length=4)
tgram_gen = tgram_step.process_track_online(test_file, 2048, 1024, chunk_size=5)
chunks = list(tgram_gen)
tgrams = np.vstack(chunks)
assert np.allclose(tgrams, exp_tgram)
def sample_utterance():
file_track = tracks.FileTrack('test_file', resources.sample_wav_file('wav_1.wav'))
utterance = tracks.Utterance('test', file_track)
return utterance
def test_read_samples(benchmark):
wav_path = resources.get_test_resource_path(('wav_files', 'med_len.wav'))
track = tracks.FileTrack('idx', wav_path)
benchmark(run, track)
def test_contains_track_returns_false(self, ds):
track = tracks.FileTrack('wav-1', '/some/other/path/here')
res = ds.contains_track(track)
assert not res
