def find_queries(query_dir_path):
"""Find all queries under sws2013_dev & sws2013_eval."""
# e.g. "sws2013_dev_123.wav" or "sws2013_dev_123_01.wav" -> "sws2013_dev_123"
pattern = re.compile(r"(_[0-9]{2})?\.wav")
query2tensors = defaultdict(list)
for query_path in tqdm(list(query_dir_path.glob("*.wav")),
ncols=0,
desc="Load queries"):
query_name = pattern.sub("", query_path.name)
wav_tensor, sample_rate = apply_effects_file(
str(query_path), [["channels", "1"], ["rate", "16000"], ["norm"]])
trimmed, _ = apply_effects_tensor(
wav_tensor,
sample_rate,
[
["vad", "-T", "0.25", "-p", "0.1"],
["reverse"],
["vad", "-T", "0.25", "-p", "0.1"],
["reverse"],
],
)
wav_tensor = trimmed if trimmed.size(1) >= (sample_rate *
0.5) else wav_tensor
wav_tensor = wav_tensor.squeeze(0)
query2tensors[query_name].append(wav_tensor)
return query2tensors
def test_apply_effects(self, args):
"""`apply_effects_tensor` should return identical data as sox command"""
effects = args['effects']
num_channels = args.get("num_channels", 2)
input_sr = args.get("input_sample_rate", 8000)
output_sr = args.get("output_sample_rate")
input_path = self.get_temp_path('input.wav')
reference_path = self.get_temp_path('reference.wav')
original = get_sinusoid(frequency=800,
sample_rate=input_sr,
n_channels=num_channels,
dtype='float32')
save_wav(input_path, original, input_sr)
sox_utils.run_sox_effect(input_path,
reference_path,
effects,
output_sample_rate=output_sr)
expected, expected_sr = load_wav(reference_path)
found, sr = sox_effects.apply_effects_tensor(original, input_sr,
effects)
assert sr == expected_sr
self.assertEqual(expected, found)
def test_apply_effects_tensor(self, args):
"""`apply_effects_tensor` should not crash"""
effects = args['effects']
num_channels = args.get("num_channels", 2)
input_sr = args.get("input_sample_rate", 8000)
original = get_sinusoid(
frequency=800, sample_rate=input_sr,
n_channels=num_channels, dtype='float32')
_found, _sr = sox_effects.apply_effects_tensor(original, input_sr, effects)
def extract_embedding(audio_path, embedder):
wav, sr = torchaudio.load(audio_path) # 2D
if sr != embedder.RATE:
wav, sr = ta_sox.apply_effects_tensor(
wav, sr, [["rate", str(embedder.RATE)]])
try:
emb = embedder([wav[0].cuda().float()]).cpu().numpy()
except RuntimeError:
emb = None
return emb
def _sox_convert(
waveform: torch.FloatTensor,
sample_rate: int,
effects: List[List[str]],
) -> torch.FloatTensor:
try:
import torchaudio.sox_effects as ta_sox
except ImportError:
raise ImportError("Please install torchaudio to convert audios")
return ta_sox.apply_effects_tensor(waveform, sample_rate, effects)[0]
def _convert_to_mono(waveform: torch.FloatTensor,
sample_rate: int) -> torch.FloatTensor:
if waveform.shape[0] > 1:
try:
import torchaudio.sox_effects as ta_sox
except ImportError:
raise ImportError(
"Please install torchaudio to convert multi-channel audios")
effects = [['channels', '1']]
return ta_sox.apply_effects_tensor(waveform, sample_rate, effects)[0]
return waveform
def test_apply_no_effect(self, dtype, sample_rate, num_channels, channels_first):
"""`apply_effects_tensor` without effects should return identical data as input"""
original = get_wav_data(dtype, num_channels, channels_first=channels_first)
expected = original.clone()
found, output_sample_rate = sox_effects.apply_effects_tensor(
expected, sample_rate, [], channels_first)
assert output_sample_rate == sample_rate
# SoxEffect should not alter the input Tensor object
self.assertEqual(original, expected)
# SoxEffect should not return the same Tensor object
assert expected is not found
# Returned Tensor should equal to the input Tensor
self.assertEqual(expected, found)
def augment(self, x):
speed = self.rng.uniform(1 - self.speed_range, 1 + self.speed_range)
pitch = int(self.rng.uniform(-self.pitch_range, self.pitch_range))
effects = [
['gain', '-n'],
['pitch', f"{pitch}"],
['speed', f'{speed:.2f}'],
]
x, _ = apply_effects_tensor(x,
self.sample_rate,
effects,
channels_first=True)
return x
def crop_segment(tensor, tgt_dur, sample_rate=16000):
src_dur = len(tensor) / sample_rate
random_shift = random.uniform(0, src_dur - tgt_dur)
audio_tensor, _ = apply_effects_tensor(
tensor.unsqueeze(0),
sample_rate,
[
["pad", f"{tgt_dur}", f"{tgt_dur}"],
[
"trim",
f"{tgt_dur + random_shift}",
f"{tgt_dur}",
],
],
)
return audio_tensor.squeeze(0)
def convert_waveform(
waveform: Union[np.ndarray, torch.Tensor],
sample_rate: int,
normalize_volume: bool = False,
to_mono: bool = False,
to_sample_rate: Optional[int] = None
) -> Tuple[Union[np.ndarray, torch.Tensor], int]:
"""convert a waveform:
- to a target sample rate
- from multi-channel to mono channel
- volume normalization
Args:
waveform (numpy.ndarray or torch.Tensor): 2D original waveform
(channels x length)
sample_rate (int): original sample rate
normalize_volume (bool): perform volume normalization
to_mono (bool): convert to mono channel if having multiple channels
to_sample_rate (Optional[int]): target sample rate
Returns:
waveform (numpy.ndarray): converted 2D waveform (channels x length)
sample_rate (float): target sample rate
"""
try:
import torchaudio.sox_effects as ta_sox
except ImportError:
raise ImportError("Please install torchaudio: pip install torchaudio")
effects = []
if normalize_volume:
effects.append(["gain", "-n"])
if to_sample_rate is not None and to_sample_rate != sample_rate:
effects.append(["rate", f"{to_sample_rate}"])
if to_mono and waveform.shape[0] > 1:
effects.append(["channels", "1"])
if len(effects) > 0:
is_np_input = isinstance(waveform, np.ndarray)
_waveform = torch.from_numpy(waveform) if is_np_input else waveform
converted, converted_sample_rate = ta_sox.apply_effects_tensor(
_waveform, sample_rate, effects)
if is_np_input:
converted = converted.numpy()
return converted, converted_sample_rate
return waveform, sample_rate
def test_apply_effects_tensor(self, args):
effects = args['effects']
channels_first = True
num_channels = args.get("num_channels", 2)
input_sr = args.get("input_sample_rate", 8000)
trans = SoxEffectTensorTransform(effects, input_sr, channels_first)
path = self.get_temp_path('sox_effect.zip')
torch.jit.script(trans).save(path)
trans = torch.jit.load(path)
wav = get_sinusoid(
frequency=800, sample_rate=input_sr,
n_channels=num_channels, dtype='float32', channels_first=channels_first)
found, sr_found = trans(wav)
expected, sr_expected = sox_effects.apply_effects_tensor(
wav, input_sr, effects, channels_first)
assert sr_found == sr_expected
self.assertEqual(expected, found)
def _get_torchaudio_fbank(waveform,
sample_rate,
n_bins=80) -> Optional[np.ndarray]:
"""Get mel-filter bank features via TorchAudio."""
try:
import torch
import torchaudio.compliance.kaldi as ta_kaldi
import torchaudio.sox_effects as ta_sox
waveform = torch.from_numpy(waveform)
if len(waveform.shape) == 1:
# Mono channel: D -> 1 x D
waveform = waveform.unsqueeze(0)
else:
# Merge multiple channels to one: C x D -> 1 x D
waveform, _ = ta_sox.apply_effects_tensor(waveform, sample_rate,
['channels', '1'])
features = ta_kaldi.fbank(waveform,
num_mel_bins=n_bins,
sample_frequency=sample_rate)
return features.numpy()
except ImportError:
return None
def process_wav(wav_files, out_dir, tempo, random_tempo=False):
# NOTE: must be careful about random seed
if random_tempo:
random.seed(int(tempo * 100))
for wav_file in tqdm(wav_files):
if random_tempo:
tempo = round(random.uniform(0.9, 1.1), 2)
wav, sr = librosa.load(wav_file, sr=None)
x = torch.from_numpy(wav).view(1, -1)
# pitch shift by sox
effects = [["tempo", f"{tempo}"], ["rate", f"{sr}"]]
y, y_sr = apply_effects_tensor(x, sr, effects)
y = y.view(-1)
assert y_sr == sr
postfix = str(tempo).replace("-", "minus") + "tempo_aug"
out_file = join(out_dir,
basename(wav_file).replace(".wav", f"_{postfix}.wav"))
sf.write(out_file, y.numpy(), sr)
def forward(self, tensor: torch.Tensor):
return sox_effects.apply_effects_tensor(tensor, self.sample_rate,
self.effects,
self.channels_first)
def forward(self, wav_tensor: torch.Tensor,
sample_rate: int) -> torch.Tensor:
wav_tensor, _ = apply_effects_tensor(wav_tensor, sample_rate,
self.effects)
return wav_tensor