def create_sample_dataset(temp_dir):
ds = audiomate.Corpus(str(temp_dir))
file_1_path = resources.sample_wav_file('wav_1.wav')
file_2_path = resources.sample_wav_file('wav_2.wav')
file_3_path = resources.get_resource_path(
['audio_formats', 'flac_1_16k_16b.flac'])
file_1 = ds.new_file(file_1_path, track_idx='wav_1')
file_2 = ds.new_file(file_2_path, track_idx='wav_2')
file_3 = ds.new_file(file_3_path, track_idx='wav_3')
issuer_1 = Speaker('spk-1', gender=Gender.MALE)
issuer_2 = Speaker('spk-2', gender=Gender.FEMALE)
issuer_3 = Issuer('spk-3')
ds.import_issuers([issuer_1, issuer_2, issuer_3])
# 2.5951875
utt_1 = ds.new_utterance('utt-1', file_1.idx, issuer_idx=issuer_1.idx)
utt_2 = ds.new_utterance('utt-2',
file_2.idx,
issuer_idx=issuer_2.idx,
start=0,
end=1.5)
utt_3 = ds.new_utterance('utt-3',
file_2.idx,
issuer_idx=issuer_2.idx,
start=1.5,
end=2.5)
# 5.0416875
utt_4 = ds.new_utterance('utt-4', file_3.idx, issuer_idx=issuer_3.idx)
utt_1.set_label_list(
LabelList(audiomate.corpus.LL_WORD_TRANSCRIPT,
labels=[Label('who am i')]))
utt_2.set_label_list(
LabelList(audiomate.corpus.LL_WORD_TRANSCRIPT,
labels=[Label('who are you')]))
utt_3.set_label_list(
LabelList(audiomate.corpus.LL_WORD_TRANSCRIPT,
labels=[Label('who is he')]))
utt_4.set_label_list(
LabelList(audiomate.corpus.LL_WORD_TRANSCRIPT,
labels=[Label('who are they')]))
train_filter = subview.MatchingUtteranceIdxFilter(
utterance_idxs={'utt-1', 'utt-2', 'utt-3'})
sv_train = subview.Subview(ds, filter_criteria=[train_filter])
dev_filter = subview.MatchingUtteranceIdxFilter(utterance_idxs={'utt-4'})
sv_dev = subview.Subview(ds, filter_criteria=[dev_filter])
ds.import_subview('train', sv_train)
ds.import_subview('dev', sv_dev)
return ds
def test_validate_passes(self):
corpus = audiomate.Corpus()
corpus.new_file(resources.sample_wav_file('wav_1.wav'), 'wav1')
corpus.new_file(resources.sample_wav_file('wav_2.wav'), 'wav2')
val = validation.TrackReadValidator()
res = val.validate(corpus)
assert res.passed
def test_validate_doesnt_pass(self):
corpus = audiomate.Corpus()
corpus.new_file(resources.sample_wav_file('wav_1.wav'), 'wav1')
corpus.new_file(resources.sample_wav_file('invalid_audio.wav'), 'wav2')
val = validation.TrackReadValidator()
res = val.validate(corpus)
assert not res.passed
assert len(res.invalid_items) == 1
assert 'wav2' in res.invalid_items
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
tgram_step = pipeline.Tempogram(win_length=4)
tgram_gen = tgram_step.process_file_online(test_file_path,
2048,
1024,
chunk_size=5)
chunks = list(tgram_gen)
tgrams = np.vstack(chunks)
assert np.allclose(tgrams, exp_tgram)
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
onset = pipeline.OnsetStrength()
onset_gen = onset.process_file_online(test_file_path,
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_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 ds():
ds = resources.create_dataset()
file_1_path = resources.sample_wav_file('wav_1.wav')
file_2_path = resources.get_resource_path(
('audio_formats', 'mp3_2_44_1k_16b.mp3'))
file_3_path = resources.get_resource_path(
('audio_formats', 'flac_1_16k_16b.flac'))
file_4_path = resources.sample_wav_file('wav_4.wav')
ds.tracks['wav-1'].path = file_1_path
ds.tracks['wav_2'].path = file_2_path
ds.tracks['wav_3'].path = file_3_path
ds.tracks['wav_4'].path = file_4_path
return ds
def test_compute_cleanup_after_one_utterance(self):
test_file_path = resources.sample_wav_file('wav_1.wav')
y, sr = librosa.load(test_file_path, sr=None)
frames = librosa.util.frame(y, frame_length=2048, hop_length=1024).T
# EXPECTED
S = np.abs(librosa.stft(y, 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=11,
center=True).T
# ACTUAL
tgram_step = pipeline.Tempogram(win_length=11)
# FIRST RUN
tgrams = tgram_step.process_frames(frames, sr, last=True)
assert np.allclose(tgrams, exp_tgram)
# SECOND RUN
tgrams = tgram_step.process_frames(frames, sr, last=True)
assert np.allclose(tgrams, exp_tgram)
def test_read_samples(self):
file = assets.File('wav', resources.sample_wav_file('wav_1.wav'))
issuer = assets.Issuer('toni')
utt = assets.Utterance('test',
file,
issuer=issuer,
start=1.0,
end=2.30)
l1 = assets.Label('a', 0.15, 0.448)
l2 = assets.Label('a', 0.5, 0.73)
ll = assets.LabelList(labels=[l1, l2])
utt.set_label_list(ll)
expected, __ = librosa.core.load(file.path,
sr=None,
offset=1.15,
duration=0.298)
assert np.array_equal(l1.read_samples(), expected)
expected, __ = librosa.core.load(file.path,
sr=None,
offset=1.5,
duration=0.23)
assert np.array_equal(l2.read_samples(), expected)
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 setUp(self):
file = assets.File('wav', resources.sample_wav_file('wav_1.wav'))
utt = assets.Utterance('utt', file, start=0.3, end=-1)
ll = assets.LabelList()
self.test_label = assets.Label('a', start=0.5, end=-1)
ll.append(self.test_label)
utt.set_label_list(ll)
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 test_save_file_tracks(self, writer, sample_corpus, tmpdir):
# make sure relative path changes in contrast to self.ds.path
out_path = os.path.join(tmpdir.strpath, 'somesubdir')
os.makedirs(out_path)
writer.save(sample_corpus, out_path)
file_1_path = os.path.relpath(resources.sample_wav_file('wav_1.wav'),
out_path)
file_2_path = os.path.relpath(resources.sample_wav_file('wav_2.wav'),
out_path)
file_3_path = os.path.relpath(resources.sample_wav_file('wav_3.wav'),
out_path)
file_4_path = os.path.relpath(resources.sample_wav_file('wav_4.wav'),
out_path)
with open(os.path.join(out_path, 'files.txt'), 'r') as f:
file_content = f.read()
assert file_content.strip(
) == 'wav-1 {}\nwav_2 {}\nwav_3 {}\nwav_4 {}'.format(
file_1_path, file_2_path, file_3_path, file_4_path)
def test_convert_files(self, tmp_path):
source_path = resources.sample_wav_file('wav_1.wav')
target_path = tmp_path / 'out.wav'
files = [(source_path, 0, float('inf'), str(target_path))]
c = conversion.WavAudioFileConverter()
c._convert_files(files)
samples, sr = librosa.core.load(source_path, sr=None)
stored_samples, stored_sr = librosa.core.load(str(target_path), sr=None)
assert target_path.is_file()
assert stored_sr == sr
assert np.array_equal(stored_samples, samples)
def test_encode_utterance_takes_lower_index_first(self):
file = assets.File('file-idx', resources.sample_wav_file('wav_1.wav'))
utt = assets.Utterance('utt-idx', file, start=0, end=5)
ll = assets.LabelList(
labels=[assets.Label('music', 0, 3),
assets.Label('speech', 3, 5)])
utt.set_label_list(ll)
enc = label_encoding.FrameOrdinalEncoder(
['speech', 'music', 'noise'],
frame_settings=units.FrameSettings(32000, 16000),
sr=16000)
actual = enc.encode(utt)
expected = np.array([1, 1, 0, 0]).astype(np.int)
assert np.array_equal(expected, actual)
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_store_samples_sr_24(self, tmp_path):
source_path = resources.sample_wav_file('wav_1.wav')
target_path = tmp_path / 'out.wav'
files = [(source_path, 0, float('inf'), str(target_path))]
c = conversion.WavAudioFileConverter(sampling_rate=24000)
c._convert_files(files)
samples, sr = librosa.core.load(source_path, sr=24000)
stored_samples, stored_sr = librosa.core.load(str(target_path), sr=None)
assert target_path.is_file()
assert stored_sr == sr
# Don't take it too exactly
# With sox 14.4.1 it isn't that precise, expecially the first sample
assert np.allclose(stored_samples[1:], samples[1:], atol=0.001)
def test_compute(self):
test_file_path = resources.sample_wav_file('wav_1.wav')
y, sr = librosa.load(test_file_path, sr=None)
frames = librosa.util.frame(y, frame_length=2048, hop_length=1024).T
# EXPECTED
S = np.abs(librosa.stft(y, 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
onset = pipeline.OnsetStrength()
onsets = onset.process_frames(frames, sr, last=True)
assert np.allclose(onsets, exp_onsets)
def test_encode_utterance_takes_larger_label(self):
file = tracks.FileTrack('file-idx',
resources.sample_wav_file('wav_1.wav'))
utt = tracks.Utterance('utt-idx', file, start=0, end=8)
ll = annotations.LabelList(labels=[
annotations.Label('music', 0, 4.5),
annotations.Label('speech', 4.5, 8)
])
utt.set_label_list(ll)
enc = encoding.FrameOrdinalEncoder(['music', 'speech', 'noise'],
'default',
frame_settings=units.FrameSettings(
32000, 16000),
sr=16000)
actual = enc.encode_utterance(utt)
expected = np.array([0, 0, 0, 0, 1, 1, 1]).astype(np.int)
assert np.array_equal(expected, actual)
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 sample_utterance():
file = assets.File('test_file', resources.sample_wav_file('wav_1.wav'))
utterance = assets.Utterance('test', file)
return utterance
