def testFloatWavDataToSamples(self):
y = audio_io.wav_data_to_samples(self.wav_data, sample_rate=16000)
wav_io = io.BytesIO()
scipy.io.wavfile.write(wav_io, 16000, y)
y_from_float = audio_io.wav_data_to_samples(wav_io.getvalue(),
sample_rate=16000)
np.testing.assert_array_equal(y, y_from_float)
def testWavDataToSamples(self):
w = wave.open(self.wav_filename, 'rb')
w_mono = wave.open(self.wav_filename_mono, 'rb')
# Check content size.
y = audio_io.wav_data_to_samples(self.wav_data, sample_rate=16000)
y_mono = audio_io.wav_data_to_samples(self.wav_data_mono, sample_rate=22050)
self.assertEqual(
round(16000.0 * w.getnframes() / w.getframerate()), y.shape[0])
self.assertEqual(
round(22050.0 * w_mono.getnframes() / w_mono.getframerate()),
y_mono.shape[0])
# Check a few obvious failure modes.
self.assertLess(0.01, y.std())
self.assertLess(0.01, y_mono.std())
self.assertGreater(-0.1, y.min())
self.assertGreater(-0.1, y_mono.min())
self.assertLess(0.1, y.max())
self.assertLess(0.1, y_mono.max())
def _wav_to_mel(wav_audio, hparams):
"""Transforms the contents of a wav file into a series of mel spec frames."""
y = audio_io.wav_data_to_samples(wav_audio, hparams.sample_rate)
mel = librosa.feature.melspectrogram(y,
hparams.sample_rate,
hop_length=hparams.spec_hop_length,
fmin=hparams.spec_fmin,
n_mels=hparams.spec_n_bins,
htk=hparams.spec_mel_htk).astype(
np.float32)
# Transpose so that the data is in [frame, bins] format.
mel = mel.T
return mel
def _wav_to_cqt(wav_audio, hparams):
"""Transforms the contents of a wav file into a series of CQT frames."""
y = audio_io.wav_data_to_samples(wav_audio, hparams.sample_rate)
cqt = np.abs(librosa.core.cqt(y,
hparams.sample_rate,
hop_length=hparams.spec_hop_length,
fmin=hparams.spec_fmin,
n_bins=hparams.spec_n_bins,
bins_per_octave=hparams.cqt_bins_per_octave),
dtype=np.float32)
# Transpose so that the data is in [frame, bins] format.
cqt = cqt.T
return cqt
def _wav_to_framed_samples(wav_audio, hparams):
"""Transforms the contents of a wav file into a series of framed samples."""
y = audio_io.wav_data_to_samples(wav_audio, hparams.sample_rate)
hl = hparams.spec_hop_length
n_frames = int(np.ceil(y.shape[0] / hl))
frames = np.zeros((n_frames, hl), dtype=np.float32)
# Fill in everything but the last frame which may not be the full length
cutoff = (n_frames - 1) * hl
frames[:n_frames - 1, :] = np.reshape(y[:cutoff], (n_frames - 1, hl))
# Fill the last frame
remain_len = len(y[cutoff:])
frames[n_frames - 1, :remain_len] = y[cutoff:]
return frames
def mix_examples(mixid_exs, sample_rate, load_audio_with_librosa):
"""Mix several Examples together to create a new example."""
mixid, exs = mixid_exs
del mixid
example_samples = []
example_sequences = []
for ex_str in exs:
ex = tf.train.Example.FromString(ex_str)
wav_data = ex.features.feature['audio'].bytes_list.value[0]
if load_audio_with_librosa:
samples = audio_io.wav_data_to_samples_librosa(
wav_data, sample_rate)
else:
samples = audio_io.wav_data_to_samples(wav_data, sample_rate)
example_samples.append(samples)
ns = music_pb2.NoteSequence.FromString(
ex.features.feature['sequence'].bytes_list.value[0])
example_sequences.append(ns)
mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences(
individual_samples=example_samples,
sample_rate=sample_rate,
individual_sequences=example_sequences)
mixed_wav_data = audio_io.samples_to_wav_data(mixed_samples, sample_rate)
mixed_id = '::'.join(['mixed'] + [ns.id for ns in example_sequences])
mixed_sequence.id = mixed_id
mixed_filename = '::'.join(['mixed'] +
[ns.filename for ns in example_sequences])
mixed_sequence.filename = mixed_filename
examples = list(
audio_label_data_utils.process_record(mixed_wav_data,
mixed_sequence,
mixed_id,
min_length=0,
max_length=-1,
sample_rate=sample_rate))
assert len(examples) == 1
return examples[0]
def testSplitAudioLabelData(self):
wav_data, sequence = self._CreateSyntheticExample()
records = audio_label_data_utils.process_record(
wav_data, sequence, 'test', sample_rate=SAMPLE_RATE)
for record in records:
audio = record.features.feature['audio'].bytes_list.value[0]
velocity_range = music_pb2.VelocityRange.FromString(
record.features.feature['velocity_range'].bytes_list.value[0])
note_sequence = music_pb2.NoteSequence.FromString(
record.features.feature['sequence'].bytes_list.value[0])
expected_samples = np.zeros(10 * SAMPLE_RATE)
np.testing.assert_array_equal(
expected_samples,
audio_io.wav_data_to_samples(audio, sample_rate=SAMPLE_RATE))
self.assertEqual(velocity_range.min, 20)
self.assertEqual(velocity_range.max, 80)
self.assertEqual(note_sequence.notes[0].velocity, 20)
self.assertEqual(note_sequence.notes[0].end_time, 5.)
self.assertEqual(note_sequence.notes[1].velocity, 80)
self.assertEqual(note_sequence.notes[1].end_time, 10.)
def process_record(wav_data,
ns,
example_id,
min_length=5,
max_length=20,
sample_rate=16000,
allow_empty_notesequence=False,
load_audio_with_librosa=False):
"""Split a record into chunks and create an example proto.
To use the full length audio and notesequence, set min_length=0 and
max_length=-1.
Args:
wav_data: audio data in WAV format.
ns: corresponding NoteSequence.
example_id: id for the example proto
min_length: minimum length in seconds for audio chunks.
max_length: maximum length in seconds for audio chunks.
sample_rate: desired audio sample rate.
allow_empty_notesequence: whether an empty NoteSequence is allowed.
load_audio_with_librosa: Use librosa for sampling. Works with 24-bit wavs.
Yields:
Example protos.
"""
try:
if load_audio_with_librosa:
samples = audio_io.wav_data_to_samples_librosa(wav_data, sample_rate)
else:
samples = audio_io.wav_data_to_samples(wav_data, sample_rate)
except audio_io.AudioIOReadError as e:
print('Exception %s', e)
return
samples = librosa.util.normalize(samples, norm=np.inf)
# Add padding to samples if notesequence is longer.
pad_to_samples = int(math.ceil(ns.total_time * sample_rate))
padding_needed = pad_to_samples - samples.shape[0]
if padding_needed > 5 * sample_rate:
raise ValueError(
'Would have padded {} more than 5 seconds to match note sequence total '
'time. ({} original samples, {} sample rate, {} sample seconds, '
'{} sequence seconds) This likely indicates a problem with the source '
'data.'.format(
example_id, samples.shape[0], sample_rate,
samples.shape[0] / sample_rate, ns.total_time))
samples = np.pad(samples, (0, max(0, padding_needed)), 'constant')
if max_length == min_length:
splits = np.arange(0, ns.total_time, max_length)
elif max_length > 0:
splits = find_split_points(ns, samples, sample_rate, min_length, max_length)
else:
splits = [0, ns.total_time]
velocity_range = velocity_range_from_sequence(ns)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < min_length:
continue
if start == 0 and end == ns.total_time:
new_ns = ns
else:
new_ns = sequences_lib.extract_subsequence(ns, start, end)
if not new_ns.notes and not allow_empty_notesequence:
tf.logging.warning('skipping empty sequence')
continue
if start == 0 and end == ns.total_time:
new_samples = samples
else:
# the resampling that happen in crop_wav_data is really slow
# and we've already done it once, avoid doing it twice
new_samples = audio_io.crop_samples(samples, sample_rate, start,
end - start)
new_wav_data = audio_io.samples_to_wav_data(new_samples, sample_rate)
yield create_example(
example_id, new_ns, new_wav_data, velocity_range=velocity_range)
