def test_read_wavfile(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
self.assertEqual(16000, sample_rate)
self.assertEqual(len(wav), 14118)
# 16-bit PCM
self.assertGreaterEqual(min(wav), -32768)
self.assertLessEqual(max(wav), 32768)
def test_preprocess_wav(self):
_, wav = read_wavfile(self.wav_filepath)
wav_distorted = preprocess_wav(wav, distort=True)
wav_preprocessed = preprocess_wav(wav, distort=False)
self.assertEqual(16000, len(wav_distorted))
self.assertEqual(16000, len(wav_preprocessed))
self.assertGreaterEqual(1, np.abs(wav_distorted).max())
self.assertGreaterEqual(1, np.abs(wav_preprocessed).max())
self.assertLess(500, len(set(wav_preprocessed)))
self.assertLess(500, len(set(wav_distorted)))
def load_random_real_noise_clip(data_version: str) -> np.array:
"""
Loads a random noise clip from the _background_noise_ folder and returns it
:param data_version: specifies the version of the data to use (str {"0.01", "0.02"})
:return: real noise clip (np.array)
"""
path = os.path.join(get_training_data_path(data_version=data_version), "_background_noise_")
filename = random.choice(list(filter(lambda x: x.endswith(".wav"), os.listdir(path))))
_, clip = read_wavfile(os.path.join(path, filename))
return clip
def test_resample_wavfile(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = resample_wavfile(wav=wav, factor=1)
self.assertEqual(len(wav), len(wav_processed))
self.assertAlmostEqual(wav.sum(), wav_processed.sum(), delta=0.00001)
wav_processed = resample_wavfile(wav=wav, factor=2)
self.assertEqual(len(wav), len(wav_processed))
self.assertLess(np.std(wav), np.std(wav_processed))
wav_processed = resample_wavfile(wav=wav, factor=0.5)
self.assertEqual(len(wav), len(wav_processed))
self.assertGreater(np.std(wav), np.std(wav_processed))
def load_real_noise_clips(data_version: str) -> np.array:
"""
Loads all the available real noise clips, which are located under the _background_noise_ folder
:param data_version: specifies the version of the data to use (str {"0.01", "0.02"})
:return: list of real noise clips (list of np.array)
"""
clips = []
path = os.path.join(get_training_data_path(data_version=data_version), "_background_noise_")
for filename in filter(lambda x: x.endswith(".wav"), os.listdir(path)):
_, wav = read_wavfile(os.path.join(path, filename))
clips.append(wav)
return clips
def test_saturate_wavfile(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav = normalize_wavfile(wav, normalize_to_peak=True)
wav_processed = saturate_wavfile(wav, 1)
self.assertListEqual(wav.tolist(), wav_processed.tolist())
wav_processed = saturate_wavfile(wav, 2) # Atenuate
self.assertAlmostEqual(wav.mean(), wav_processed.mean(), delta=0.05)
self.assertGreater(np.std(wav), np.std(wav_processed))
self.assertGreater(max(abs(wav)), max(abs(wav_processed)))
wav_processed = saturate_wavfile(wav, 0.2) # Saturate
self.assertAlmostEqual(wav.mean(), wav_processed.mean(), delta=0.05)
self.assertLess(np.std(wav), np.std(wav_processed))
self.assertEqual(max(abs(wav)), max(abs(wav_processed)))
def test_generate_augmented_wav(self):
filepath_augmented = os.path.join(
get_augmented_data_path(data_version=self.data_version),
"testfolder", "examples", "testaudio.wav")
if os.path.exists(filepath_augmented):
os.remove(filepath_augmented)
generate_augmented_wav(data_version=self.data_version,
filepath=self.wav_filepath,
folder="testfolder")
self.assertTrue(os.path.exists(filepath_augmented))
_, wav = read_wavfile(filepath_augmented)
self.assertEqual(16000, len(wav))
self.assertGreaterEqual(1, np.abs(wav).max())
os.remove(filepath_augmented)
shutil.rmtree(os.path.split(os.path.split(filepath_augmented)[0])[0])
def test_fix_wavfile_length(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = fix_wavfile_length(wav, 8000) # Cut
self.assertEqual(8000, len(wav_processed))
self.assertAlmostEqual(wav_processed.mean(), wav.mean(), delta=0.05)
wav_processed = fix_wavfile_length(wav, 30000) # Pad
self.assertEqual(30000, len(wav_processed))
self.assertEqual(wav_processed.sum(), wav.sum())
self.assertListEqual([0] * 1000, wav_processed[:1000].tolist())
self.assertListEqual([0] * 1000, wav_processed[-1000:].tolist())
wav_processed = fix_wavfile_length(wav, 30001) # Pad odd
self.assertEqual(30001, len(wav_processed))
self.assertEqual(wav_processed.sum(), wav.sum())
self.assertListEqual([0] * 1000, wav_processed[:1000].tolist())
self.assertListEqual([0] * 1000, wav_processed[-1000:].tolist())
def test_fix_wavfile_duration(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = fix_wavfile_duration(wav=wav,
sample_rate=sample_rate,
duration=len(wav) / sample_rate)
self.assertListEqual(list(wav), list(wav_processed))
wav_processed = fix_wavfile_duration(wav=wav,
sample_rate=sample_rate,
duration=0.3)
self.assertEqual(len(wav_processed), int(0.3 * sample_rate))
wav_processed = fix_wavfile_duration(wav=wav,
sample_rate=sample_rate,
duration=3)
self.assertEqual(len(wav_processed), int(3 * sample_rate))
self.assertListEqual([0] * 16000, list(wav_processed[:16000]))
self.assertListEqual([0] * 16000, list(wav_processed[-16000:]))
self.assertEqual(np.sum(wav), np.sum(wav_processed))
def generate_augmented_wav(data_version: str, filepath: str, folder: str) -> None:
"""
Given a filepath of a wav file, loads it, preprocesses it and stores it in the default folder for augmentations.
:param data_version: specifies the version of the data to use (str {"0.01", "0.02"})
:param filepath: file path of an existing wav file (str)
:param folder: Name of the folder which will contain this version of the augmentation (str)
:return: None (void)
"""
try:
folder_class = os.path.split(os.path.split(filepath)[-2])[-1]
output_path = os.path.join(get_augmented_data_folder(data_version=data_version, folder=folder), folder_class)
if not os.path.exists(output_path):
os.makedirs(output_path)
sample_rate, wav = read_wavfile(filepath)
wav = preprocess_wav(wav, distort=True)
# Add suffix
filename = os.path.split(filepath)[-1]
output_filepath = os.path.join(output_path, filename)
wavfile.write(output_filepath, sample_rate, wav)
except:
print("Error found with file {}".format(filepath))
def test_time_offset(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = time_offset_wavfile(wav=wav, shift_factor=0)
self.assertEqual(len(wav), len(wav_processed))
self.assertAlmostEqual(wav.sum(), wav_processed.sum(), delta=0.00001)
wav_processed = time_offset_wavfile(wav=wav,
shift_factor=1) # Full_shift
self.assertEqual(np.abs(wav_processed).sum(), 0)
wav_processed = time_offset_wavfile(
wav=wav, shift_factor=-1) # Full shift inverse
self.assertEqual(np.abs(wav_processed).sum(), 0)
wav_processed = time_offset_wavfile(wav=wav,
shift_factor=0.5) # Half shift
self.assertEqual(len(wav), len(wav_processed))
self.assertListEqual(wav.tolist()[:int(len(wav) / 2)],
wav_processed.tolist()[-int(len(wav) / 2):])
wav_processed = time_offset_wavfile(
wav=wav, shift_factor=-0.5) # Half shift inverse
self.assertEqual(len(wav), len(wav_processed))
self.assertListEqual(wav.tolist()[-int(len(wav) / 2):],
wav_processed.tolist()[:int(len(wav) / 2)])
def load_data(self, file_paths: list, add_noise: bool, load_targets: bool = True) -> None:
self.audios = []
if load_targets:
self.targets = []
# Load data
for file_path in tqdm(file_paths):
if os.path.exists(file_path):
try:
_, wav = read_wavfile(file_path)
wav = preprocess_wav(wav, distort=False)
self.audios.append(wav)
if load_targets:
target = os.path.split(os.path.split(file_path)[0])[1]
self.targets.append(target)
except:
self.corrupted_file_paths.append(file_path)
print(f"Error reading {file_path}")
else:
self.corrupted_file_paths.append(file_path)
print(f"Fatal error, file {file_path} not found")
if add_noise:
n_artificial_noise_samples = int(0.05 * len(self.audios))
n_real_noise_samples = int(0.15 * len(self.audios))
n_empty_samples = int(0.0025 * len(self.audios))
for i in tqdm(range(n_real_noise_samples)):
wav = get_random_real_noise_subclip(data_version=self.data_version, n_samples=16000,
noise_clips=self.noise_clips)
wav = preprocess_wav(wav, distort=False)
self.audios.append(wav)
if load_targets: self.targets.append("silence")
for i in range(n_artificial_noise_samples):
wav = generate_white_noise_clip(16000)
wav = preprocess_wav(wav, distort=False)
self.audios.append(wav)
if load_targets: self.targets.append("silence")
for i in range(n_empty_samples):
self.audios.append(np.zeros_like(wav).astype(np.float32))
if load_targets: self.targets.append("silence")
def test_pitch_shift(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = pitch_shift_wavfile(wav=wav,
sr=sample_rate,
n_octaves=0)
self.assertEqual(len(wav), len(wav_processed))
self.assertAlmostEqual(wav.sum(), wav_processed.sum(), delta=0.00001)
wav_processed = pitch_shift_wavfile(wav=wav,
sr=sample_rate,
n_octaves=0.5)
self.assertEqual(len(wav), len(wav_processed))
# Check fundamental armonic movement
fft_wav = np.fft.fft(wav)
fft_wav = fft_wav[:int(len(fft_wav) / 2)]
fft_wav = np.abs(fft_wav)
fft_wav_proc = np.fft.fft(wav_processed)
fft_wav_proc = fft_wav_proc[:int(len(fft_wav_proc) / 2)]
fft_wav_proc = np.abs(fft_wav_proc)
self.assertLess(np.argmax(fft_wav), np.argmax(fft_wav_proc))
self.assertEqual(np.max(np.abs(wav)),
np.max(np.abs(wav_processed))) # Check peak
wav_processed = pitch_shift_wavfile(wav=wav,
sr=sample_rate,
n_octaves=-0.5)
self.assertEqual(len(wav), len(wav_processed))
# Check fundamental armonic movement
fft_wav = np.fft.fft(wav)
fft_wav = fft_wav[:int(len(fft_wav) / 2)]
fft_wav = np.abs(fft_wav)
fft_wav_proc = np.fft.fft(wav_processed)
fft_wav_proc = fft_wav_proc[:int(len(fft_wav_proc) / 2)]
fft_wav_proc = np.abs(fft_wav_proc)
self.assertGreater(np.argmax(fft_wav), np.argmax(fft_wav_proc))
self.assertEqual(np.max(np.abs(wav)),
np.max(np.abs(wav_processed))) # Check peak
def test_add_noise(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = add_noise_to_wavfile(wav=wav,
amplitude_factor=0,
clip_to_original_range=False)
self.assertEqual(len(wav), len(wav_processed))
self.assertAlmostEqual(wav.sum(), wav_processed.sum(), delta=0.00001)
wav_processed = add_noise_to_wavfile(wav=wav,
amplitude_factor=0,
clip_to_original_range=True)
self.assertEqual(len(wav), len(wav_processed))
self.assertAlmostEqual(wav.sum(), wav_processed.sum(), delta=0.00001)
wav_processed = add_noise_to_wavfile(wav=wav,
amplitude_factor=0.5,
clip_to_original_range=False)
self.assertEqual(len(wav), len(wav_processed))
self.assertNotEquals(np.abs(wav).max(), np.abs(wav_processed).max())
self.assertLess(np.std(wav), np.std(wav_processed))
wav_processed = add_noise_to_wavfile(wav=wav,
amplitude_factor=0.5,
clip_to_original_range=True)
self.assertEqual(len(wav), len(wav_processed))
self.assertEquals(np.abs(wav).max(), np.abs(wav_processed).max())
def test_normalize_wavfile_formats(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_0 = wav / 32768
wav_1 = wav_0 * 2147483648
wav_2 = wav
wav_3 = 255 * (wav_0 + 1) / 2
wav_0_processed = normalize_wavfile(wav_0, 0)
wav_1_processed = normalize_wavfile(wav_1, 1)
wav_2_processed = normalize_wavfile(wav_2, 2)
wav_3_processed = normalize_wavfile(wav_3, 3)
wavs_peak_processed = [
normalize_wavfile(wav, normalize_to_peak=True)
for wav in [wav_0, wav_1, wav_2, wav_3]
]
self.assertTrue(
max(wav_0_processed) == max(wav_1_processed) == max(
wav_2_processed) == max(wav_3_processed))
self.assertTrue(
min(wav_0_processed) == min(wav_1_processed) == min(
wav_2_processed) == min(wav_3_processed))
self.assertLessEqual(max(wav_0_processed), 1)
self.assertAlmostEqual(max(wav_0_processed), 1, delta=0.75)
self.assertGreaterEqual(min(wav_0_processed), -1)
self.assertAlmostEqual(min(wav_0_processed), -1, delta=0.75)
for wav_processed in wavs_peak_processed:
self.assertGreaterEqual(1, wav_processed.max())
self.assertLessEqual(-1, wav_processed.max())
wav_processed = normalize_wavfile(np.zeros(16000),
normalize_to_peak=True)
self.assertEqual(0, np.abs(wav_processed).sum())
with self.assertRaises(ValueError):
wav_corrupted = np.random.rand(16000)
wav_corrupted[4000] = np.nan
normalize_wavfile(wav_corrupted, normalize_to_peak=True)
with self.assertRaises(ValueError):
normalize_wavfile(np.ones(16000) * np.nan, normalize_to_peak=True)
def test_normalize_wavfile_to_peaks(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = normalize_wavfile(wav, normalize_to_peak=True)
self.assertGreaterEqual(min(wav_processed), -1)
self.assertLessEqual(max(wav_processed), 1)
self.assertEqual(max(max(wav_processed), -min(wav_processed)), 1)
def test_draw_random_subclip(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
subclip = draw_random_subclip(wav, 1000)
self.assertEqual(1000, len(subclip))
self.assertLess(0.5, np.max(np.array(subclip) != 0))
def test_randomly_distort_wavfile(self):
sample_rate, wav = read_wavfile(self.wav_filepath)
wav_processed = randomly_distort_wavfile(wav=wav, sr=sample_rate)
self.assertEqual(len(wav), len(wav_processed))