def test_get_samples_from(self, timit_like_path, timit_like_datapath, timit_like_gtpath):
"""Test sampling iterator."""
gt_getter = TimitGroundTruth(timit_like_path, timit_like_datapath, timit_like_gtpath)
audio_extractor = WindowedAudio(1024, 512, 16000, normalize=True, padding=True)
mfcc_normalized = WindowedMFCC(1024, 512, 16000, n_mfcc=32, normalize=True)
sampler = WindowedSegmentSampler([audio_extractor], gt_getter, 10000)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert np.sum(ground_truth) == 1
assert sample.shape == (1, 18, 1024)
assert np.isclose(ground_truth[gt_getter.get_index_from("h#")], .97, atol=0.01)
assert np.isclose(ground_truth[gt_getter.get_index_from("p")], .03, atol=0.01)
sampler = WindowedSegmentSampler([audio_extractor, mfcc_normalized], gt_getter, 10000)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert sample.shape == (1, 18, 1056)
# consume all iterator
for _ in iterator:
pass
audio_extractor_no_pad = WindowedAudio(1024, 512, 16000, normalize=True, padding=False)
sampler = WindowedSegmentSampler([audio_extractor_no_pad], gt_getter, 10000)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert np.sum(ground_truth) == 1
assert sample.shape == (1, 18, 1024)
print(ground_truth)
assert np.isclose(ground_truth[gt_getter.get_index_from("h#")], 0.93, atol=0.01)
assert np.isclose(ground_truth[gt_getter.get_index_from("p")], 0.05, atol=0.01)
assert np.isclose(ground_truth[gt_getter.get_index_from("iy")], 0.02, atol=0.01)
def test_get_samples_from(self, timit_like_path, timit_like_datapath, timit_like_gtpath):
"""Test sampling iterator."""
gt_getter = TimitGroundTruth(timit_like_path, timit_like_datapath, timit_like_gtpath)
audio_extractor = WindowedAudio(1024, 512, 16000, normalize=True, padding=True)
mfcc_normalized = WindowedMFCC(1024, 512, 16000, n_mfcc=32, normalize=True)
sampler = FileSampler([audio_extractor], gt_getter)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert np.isclose(np.sum(ground_truth), 0.99, atol=0.01)
assert sample.shape == (1, 130, 1024)
sampler = FileSampler([audio_extractor, mfcc_normalized], gt_getter)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert sample.shape == (1, 130, 1056)
# consume all iterator
for _ in iterator:
pass
audio_extractor_no_pad = WindowedAudio(1024, 512, 16000, normalize=True, padding=False)
sampler = FileSampler([audio_extractor_no_pad], gt_getter, 4000)
iterator = sampler.get_samples_from("train")
sample, ground_truth = next(iterator)
assert np.isclose(np.sum(ground_truth), 0.99, atol=0.01)
assert sample.shape == (1, 128, 1024)
def test_init(self, timit_like_path, timit_like_datapath, timit_like_gtpath):
"""Test instantiation."""
gt_getter = TimitGroundTruth(timit_like_path, timit_like_datapath, timit_like_gtpath)
audio_extractor = WindowedAudio(1024, 512, 16000, normalize=True, padding=True)
mfcc_extractor = WindowedMFCC(1024, 512, 16000, n_mfcc=32, normalize=True)
try:
FileSampler([audio_extractor], gt_getter)
FileSampler([audio_extractor, mfcc_extractor], gt_getter)
except Exception as exception:
pytest.fail(f"Unexpected error: {exception}")
def test_get_output_description(self, timit_like_path, timit_like_datapath, timit_like_gtpath):
"""Test the description methods."""
gt_getter = TimitGroundTruth(timit_like_path, timit_like_datapath, timit_like_gtpath)
audio_extractor = WindowedAudio(1024, 512, 16000, normalize=True, padding=True)
mfcc_normalized = WindowedMFCC(1024, 512, 16000, n_mfcc=32, normalize=True)
sampler = WindowedSegmentSampler([audio_extractor, mfcc_normalized], gt_getter, 1536)
expected_result = {
"samples": {
"WindowedAudio": list(range(0, 1024)),
"WindowedMFCC": list(range(1024, 1056))
},
"ground_truth": PHON
}
assert sampler.get_output_description() == expected_result
def test_fill_cpu_ram(self, timit_like_path, timit_like_datapath,
timit_like_gtpath):
"""Test the fill cpu rarm functionality."""
gt_getter = TimitGroundTruth(timit_like_path, timit_like_datapath,
timit_like_gtpath)
audio_extractor = WindowedAudio(1024,
512,
16000,
normalize=True,
padding=True)
mfcc_extractor = WindowedMFCC(1024,
512,
16000,
n_mfcc=32,
normalize=True)
sampler = WindowedSegmentSampler([audio_extractor, mfcc_extractor],
gt_getter, 8000)
try:
get_dataloader_fixed_size(sampler, 32, "train")
get_dataloader_fixed_size(sampler, 16, "test")
except Exception as exception:
pytest.fail(f"Unexpected error: {exception}")
def test_windowed(self, testing_audio):
"""Test the WindowedMFCC class."""
signal, samplerate = testing_audio
try:
feature_extractor = WindowedMFCC(1024, 512, 16000, n_mfcc=32, normalize=True)
mfcc_normalized = feature_extractor.process(signal, samplerate)
feature_extractor_no = WindowedMFCC(1024, 1024, 16000, n_mfcc=32, normalize=True)
mfcc_no_overlap = feature_extractor_no.process(signal, samplerate)
except Exception as exception:
pytest.fail(f"Unexpected error: {exception}")
assert (mfcc_normalized.max() <= 1) and (mfcc_normalized.min() >= -1)
assert mfcc_normalized.shape[-1] == 32
assert mfcc_normalized.shape[-2] == 1 + int((len(signal) - 1024 + 1024) / 512)
assert mfcc_no_overlap.shape[-1] == 32
assert mfcc_no_overlap.shape[-2] == 1 + int((len(signal) - 1024 + 1024) / 1024)
from audio_loader.features.raw_audio import WindowedAudio
from audio_loader.features.mfcc import WindowedMFCC
from audio_loader.ground_truth.timit import TimitGroundTruth
from audio_loader.samplers.windowed_segments import WindowedSegmentSampler
from audio_loader.dl_frontends.pytorch.fill_ram import get_dataloader_fixed_size
timit_gt = TimitGroundTruth(
join(Path.home(), "data/darpa-timit-acousticphonetic-continuous-speech"))
print("groundtruth loaded")
raw_feature_processor = WindowedAudio(1024, 512, 16000)
raw_sampler = WindowedSegmentSampler([raw_feature_processor],
timit_gt,
8000,
overlap=0.5)
raw_train_dataloader = get_dataloader_fixed_size(raw_sampler, 32, "train")
raw_test_dataloader = get_dataloader_fixed_size(raw_sampler, 32, "test")
print("raw audio done")
mfcc_feature_processor = WindowedMFCC(1024, 512, 16000, 20)
mfcc_sampler = WindowedSegmentSampler([mfcc_feature_processor],
timit_gt,
8000,
overlap=0.5)
mfcc_train_dataloader = get_dataloader_fixed_size(mfcc_sampler, 32, "train")
mfcc_test_dataloader = get_dataloader_fixed_size(mfcc_sampler, 32, "test")
print("mfcc done")