def test_calculate_SNR_negative_1(self):
source_array = [89, -89] * 6000 + [502, -502] * 8000 + [89, -89] * 7000
source_data = reduce(lambda a, b: a + struct.pack('>h', b),
source_array[1:],
struct.pack('>h', source_array[0]))
sampling_frequency = 8000
bounds_of_speech = [(2.0 * 6000.0 / sampling_frequency,
2.0 * (6000.0 + 8000.0) / sampling_frequency)]
with self.assertRaisesRegex(AssertionError, 'Sound data are wrong!'):
vad.calculate_SNR(source_data[0:-1], sampling_frequency,
bounds_of_speech)
def test_calculate_SNR_negative_4(self):
source_array = [89, -89] * 6000 + [502, -502] * 8000 + [89, -89] * 7000
source_data = reduce(lambda a, b: a + struct.pack('>h', b),
source_array[1:],
struct.pack('>h', source_array[0]))
sampling_frequency = 8000
self.assertIsNone(
vad.calculate_SNR(source_data, sampling_frequency, []))
def test_calculate_SNR_negative_3(self):
source_array = [89, -89] * 6000 + [502, -502] * 8000 + [89, -89] * 7000
source_data = reduce(lambda a, b: a + struct.pack('>h', b),
source_array[1:],
struct.pack('>h', source_array[0]))
sampling_frequency = 8000
bounds_of_speech = [(106000.0 / sampling_frequency,
(106000.0 + 8000.0) / sampling_frequency)]
self.assertIsNone(
vad.calculate_SNR(source_data, sampling_frequency,
bounds_of_speech))
def select_sound_files_and_their_SNR(dir_name, method_name, params_of_VAD):
for cur_name in os.listdir(dir_name):
if cur_name in {'.', '..'}:
continue
full_name = os.path.join(dir_name, cur_name)
if os.path.isdir(full_name):
yield from select_sound_files_and_their_SNR(full_name, method_name, params_of_VAD)
elif cur_name.lower().endswith('.wav'):
sound, fs = load_sound(full_name)
if method_name == 'adapt':
bounds_of_speech = list(detect_spoken_frames(sound, fs, params_of_VAD))
else:
bounds_of_speech = list(detect_spoken_frames_with_webrtc(sound, fs, params_of_VAD))
snr = calculate_SNR(sound, fs, bounds_of_speech)
if snr is not None:
yield (full_name, int(round(snr)))
def test_calculate_SNR_positive_1(self):
source_array = [89, -89] * 6000 + [502, -502] * 8000 + [89, -89] * 7000
source_data = reduce(lambda a, b: a + struct.pack('>h', b),
source_array[1:],
struct.pack('>h', source_array[0]))
sampling_frequency = 8000
bounds_of_speech = [(2.0 * 6000.0 / sampling_frequency,
2.0 * (6000.0 + 8000.0) / sampling_frequency)]
silence_energy = reduce(
lambda a, b: a + b * b, source_array[0:(2 * 6000)] +
source_array[(2 * (6000 + 8000)):], vad.EPS) / (2.0 *
(6000.0 + 7000.0))
speech_energy = reduce(lambda a, b: a + b * b, source_array[
(2 * 6000):(2 * (6000 + 8000))], vad.EPS) / (2.0 * 8000.0)
target_snr = 20.0 * math.log10(speech_energy / silence_energy)
self.assertAlmostEqual(
target_snr,
vad.calculate_SNR(source_data, sampling_frequency,
bounds_of_speech))
def test_calculate_SNR_negative_2(self):
sampling_frequency = 8000
bounds_of_speech = [(2.0 * 6000.0 / sampling_frequency,
2.0 * (6000.0 + 8000.0) / sampling_frequency)]
with self.assertRaisesRegex(AssertionError, 'Sound data are wrong!'):
vad.calculate_SNR([], sampling_frequency, bounds_of_speech)