def read_audio(fp, downsample=True):
sig, sr = torchaudio.load(fp)
if downsample:
# 48khz -> 16 khz
if sig.size(0) % 3 == 0:
sig = sig[::3].contiguous()
else:
sig = sig[:-(sig.size(0) % 3):3].contiguous()
return sig, sr
def load_audio(path):
sound, _ = torchaudio.load(path)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound = sound.mean(axis=1) # multiple channels, average
return sound
def load_audio(path):
'''使用torchaudio读取音频
Args:
path(string) : 音频的路径
Returns:
sound(numpy.ndarray) : 单声道音频数据，如果是多声道进行平均(Samples * 1 channel)
'''
sound, _ = torchaudio.load(path)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound = sound.mean(axis = 1)
return sound
def load_audio(path):
'''
Input:
path : string 载入音频的路径
Output:
sound : numpy.ndarray 单声道音频数据，如果是多声道进行平均
'''
sound, _ = torchaudio.load(path)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound - sound.mean(axis=1)
return sound
def load_audio(path):
"""
Args:
path : string 载入音频的路径
Returns:
sound : numpy.ndarray 单声道音频数据，如果是多声道进行平均
"""
sound, _ = torchaudio.load(path)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound = sound.mean(axis=1)
return sound
def load_wave(path, normalize=True):
"""
Args:
path : string 载入音频的路径
Returns:
"""
sound, _ = torchaudio.load(path)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound = sound.mean(axis=1)
wave = torch.FloatTensor(sound)
if normalize:
mean = wave.mean()
std = wave.std()
wave.add_(-mean)
wave.div_(std)
return wave
def download_vctk(destination, tmp_dir=None, device="cpu"):
"""Download dataset and perform resample to 16000 Hz.
Arguments
---------
destination : str
Place to put final zipped dataset.
tmp_dir : str
Location to store temporary files. Will use `tempfile` if not provided.
device : str
Passed directly to pytorch's ``.to()`` method. Used for resampling.
"""
dataset_name = "noisy-vctk-16k"
if tmp_dir is None:
tmp_dir = tempfile.gettempdir()
final_dir = os.path.join(tmp_dir, dataset_name)
if not os.path.isdir(tmp_dir):
os.mkdir(tmp_dir)
if not os.path.isdir(final_dir):
os.mkdir(final_dir)
prefix = "https://datashare.is.ed.ac.uk/bitstream/handle/10283/2791/"
noisy_vctk_urls = [
prefix + "clean_testset_wav.zip",
prefix + "noisy_testset_wav.zip",
prefix + "testset_txt.zip",
prefix + "clean_trainset_28spk_wav.zip",
prefix + "noisy_trainset_28spk_wav.zip",
prefix + "trainset_28spk_txt.zip",
]
zip_files = []
for url in noisy_vctk_urls:
filename = os.path.join(tmp_dir, url.split("/")[-1])
zip_files.append(filename)
if not os.path.isfile(filename):
logger.info("Downloading " + url)
with urllib.request.urlopen(url) as response:
with open(filename, "wb") as tmp_file:
logger.info("... to " + tmp_file.name)
shutil.copyfileobj(response, tmp_file)
# Unzip
for zip_file in zip_files:
logger.info("Unzipping " + zip_file)
shutil.unpack_archive(zip_file, tmp_dir, "zip")
os.remove(zip_file)
# Move transcripts to final dir
shutil.move(os.path.join(tmp_dir, "testset_txt"), final_dir)
shutil.move(os.path.join(tmp_dir, "trainset_28spk_txt"), final_dir)
# Downsample
dirs = [
"noisy_testset_wav",
"clean_testset_wav",
"noisy_trainset_28spk_wav",
"clean_trainset_28spk_wav",
]
downsampler = Resample(orig_freq=48000, new_freq=16000)
for directory in dirs:
logger.info("Resampling " + directory)
dirname = os.path.join(tmp_dir, directory)
# Make directory to store downsampled files
dirname_16k = os.path.join(final_dir, directory + "_16k")
if not os.path.isdir(dirname_16k):
os.mkdir(dirname_16k)
# Load files and downsample
for filename in get_all_files(dirname, match_and=[".wav"]):
signal, rate = torchaudio.load(filename)
downsampled_signal = downsampler(signal.view(1, -1).to(device))
# Save downsampled file
torchaudio.save(
os.path.join(dirname_16k, filename[-12:]),
downsampled_signal[0].cpu(),
sample_rate=16000,
channels_first=False,
)
# Remove old file
os.remove(filename)
# Remove old directory
os.rmdir(dirname)
logger.info("Zipping " + final_dir)
final_zip = shutil.make_archive(
base_name=final_dir,
format="zip",
root_dir=os.path.dirname(final_dir),
base_dir=os.path.basename(final_dir),
)
logger.info(f"Moving {final_zip} to {destination}")
shutil.move(final_zip, os.path.join(destination, dataset_name + ".zip"))
if char == '<s>':
prev = ''
continue
hypothesis += char
prev = char
return hypothesis.replace('|', ' ')
# Load Wav2Vec2 pretrained model from Hugging Face Hub
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
# Convert the model to torchaudio format, which supports TorchScript.
model = import_huggingface_model(model)
# Remove weight normalization which is not supported by quantization.
model.encoder.transformer.pos_conv_embed.__prepare_scriptable__()
model = model.eval()
# Attach decoder
model = SpeechRecognizer(model)
# Apply quantization / script / optimize for motbile
quantized_model = torch.quantization.quantize_dynamic(
model, qconfig_spec={torch.nn.Linear}, dtype=torch.qint8)
scripted_model = torch.jit.script(quantized_model)
optimized_model = optimize_for_mobile(scripted_model)
# Sanity check
waveform , _ = torchaudio.load('scent_of_a_woman_future.wav')
print('Result:', optimized_model(waveform))
optimized_model._save_for_lite_interpreter("wav2vec2.ptl")
def test_Vol(self):
test_filepath = common_utils.get_asset_path(
'steam-train-whistle-daniel_simon.wav')
waveform, _ = torchaudio.load(test_filepath)
self._assert_consistency(T.Vol(1.1), waveform)
def _reload_signal(self):
data_signal, sample_rate = torchaudio.load(self.path)
self._sample_rate = sample_rate
self._data_signal = data_signal
def __call__(
self, file: AudioFile, sample_offset: int = 0, num_samples: int = None
) -> Tensor:
"""
Parameters
----------
file : AudioFile
Audio file.
sample_offset : int, optional
Start loading at this `sample_offset` sample. Defaults ot 0.
num_samples : int, optional
Load that many samples. Defaults to load up to the end of the file.
Returns
-------
samples : (time, channel) torch.Tensor
Samples
"""
self.is_valid(file)
original_samples = None
if isinstance(file, dict):
# file = {"samples": torch.Tensor, "sample_rate": int, [ "channel": int ]}
if "samples" in file:
original_samples = file["samples"]
original_sample_rate = file["sample_rate"]
original_total_num_samples = original_samples.shape[1]
channel = file.get("channel", None)
# file = {"audio": str or Path, [ "channel": int ]}
else:
audio_path = str(file["audio"])
(
original_total_num_samples,
original_sample_rate,
) = self.get_audio_metadata(audio_path)
channel = file.get("channel", None)
# file = str or Path
else:
audio_path = str(file)
original_total_num_samples, original_sample_rate = self.get_audio_metadata(
audio_path
)
channel = None
original_sample_offset = round(
sample_offset * original_sample_rate / self.sample_rate
)
if num_samples is None:
original_num_samples = original_total_num_samples - original_sample_offset
else:
original_num_samples = round(
num_samples * original_sample_rate / self.sample_rate
)
if original_sample_offset + original_num_samples > original_total_num_samples:
raise ValueError()
if original_samples is None:
try:
original_data, _ = torchaudio.load(
audio_path,
frame_offset=original_sample_offset,
num_frames=original_num_samples,
)
except TypeError:
raise Exception(
"It looks like you are using an unsupported version of torchaudio."
" If you have 0.6 or older, please upgrade to a newer version."
)
else:
original_data = original_samples[
:, original_sample_offset : original_sample_offset + original_num_samples
]
if channel is not None:
original_data = original_data[channel - 1 : channel, :]
result = self.downmix_and_resample(original_data, original_sample_rate)
if num_samples is not None:
# If there is an off-by-one error in the length (e.g. due to resampling), fix it.
if result.shape[-1] > num_samples:
result = result[:, :num_samples]
elif result.shape[-1] < num_samples:
diff = num_samples - result.shape[-1]
result = torch.nn.functional.pad(result, (0, diff))
return result
def load_wav(path):
signal, _ = torchaudio.load(path)
signal = signal.reshape(-1)
return signal
lines = f.readlines()
data = [(line.strip().split(',')[0], int(line.strip().split(',')[1]))
for line in lines]
model.eval()
with torch.no_grad():
all_labels = []
all_pred_labels = []
small_audios_indices = []
audio_lengths = []
wrong_sample_rate_indices = []
clean_indices = []
total_mag = []
#for batch_num, (audio, label) in enumerate(loader):
for i, (path, label) in enumerate(data):
audio, sample_rate = torchaudio.load('/storage' + path)
if sample_rate != 8000:
wrong_sample_rate_indices.append(i)
continue
if model.cnn_type == 'vgg':
min_len = 10100
else:
min_len = 12300
if audio.shape[1] < min_len:
#print("Audio at index {} is too small to pass through CNN".format(i))
small_audios_indices.append(i)
continue
if audio.shape[1] > 80000:
audio = audio[:, :80000]
clean_indices.append(i)
total_mag.append(
import torch
import torchaudio
import pytorch_lightning as pl
import matplotlib.pyplot as plt
import pandas as pd
import pathlib as Path
import os
torch.cuda.is_available()
print(os.getcwd())
csv = pd.read_csv('./data/ESC-50-master/meta/esc50.csv')
x, sr = torchaudio.load(f'./data/ESC-50-master/audio/{csv.iloc[0, 0]}')
h = torchaudio.transforms.Resample(new_freq=8000)(x)
print(h.shape)
plt.imshow(h[0])
def getDFTFeature(filepath,
win_size=1024,
win_shift=512,
preemphasis=False,
channel_first=True,
drop_dc=True,
cut_len=5160,
normalize=False):
'''
获取一个音频的对数DFT频谱
Args:
filepath: 音频路径
win_size: 窗口大小（点）
win_shift: 滑动距离（点）
preemphasis: 是否预强化。通过一阶差分弱低频强高频
channel_first: whether to put channels in first dimension
drop_dc: whether to drop DC component in spectrum (frequency==0)
cut_len: keep the fix number of points in time axis
normalize: 观察发现能量很小，对数能量在-100的数量级，有必要处理一下
Return:
(log_power_spectrum, phase_spectrum):能量谱与相位谱相叠形成的tensor
大小原本为(2C,T,M//2)，C为通道数据,T为帧数,M为FFT点数
经转置后变为(T,M//2,2C)
'''
waveform, sample_freq = torchaudio.load(filepath)
m, n = waveform.shape
# padding to 2^k
if (n - win_size) % win_shift != 0:
waveform = torch.cat(
[waveform,
torch.zeros(m, win_shift - (n - win_size) % win_shift)],
dim=1)
n = waveform.shape[1]
# split frames into rows
frame_num = (n - win_size) // win_shift + 1
strided_input = waveform.as_strided((m, frame_num, win_size),
(n, win_shift, 1))
strided_input = strided_input - torch.mean(strided_input,
dim=2).unsqueeze(2)
# pre-emphasis
preemphasis = 0.97
offset_strided_input = torch.nn.functional.pad(strided_input, (1, 0),
mode='replicate')
strided_input = strided_input - preemphasis * offset_strided_input[:, :, :
-1]
# windowed and FFT
win_func = torch.hamming_window(win_size, periodic=False)
windowed_input = strided_input * win_func
fft = torch.rfft(windowed_input, 1, normalized=False,
onesided=True) * 2 / win_size
if drop_dc:
fft = fft[:, :, 1:]
fft = fft[:, :cut_len, :]
power_spectrum = fft.pow(2).sum(3)
log_power_spectrum = torch.log10(power_spectrum) * 10
# 对于能量谱正则化处理
mean_vec = log_power_spectrum.mean(axis=1, keepdim=True)
std_vec = log_power_spectrum.std(axis=1, keepdim=True)
log_power_spectrum = (log_power_spectrum - mean_vec) / std_vec
#
phase_spectrum = fft[:, :, :, 0] / fft.pow(2).sum(3).sqrt()
phase_spectrum = torch.acos(phase_spectrum)
phase_spectrum[fft[:, :, :, 0] < 0] = -phase_spectrum[fft[:, :, :, 0] < 0]
spectrums = torch.cat([log_power_spectrum, phase_spectrum], dim=0)
if not channel_first:
spectrums = spectrums.permute(1, 2, 0)
return spectrums
def __getitem__(self, index):
np.random.seed()
sample_name = self.sample_list[index]
###############################################################
# 0. Batch settings
###############################################################
if self.sample_count % self.batch_size == 0:
self.sample_count = 0
# random global stretch & compression for the batch
if self.global_stretch:
self.frames = np.random.randint(self.min_stretch_frames,
self.max_stretch_frames + 1)
self.sample_count += 1
target = np.linspace(
-1, 1,
int(self.cfg["sample_duration"] * self.cfg["vid_framerate"]))
###############################################################
# 1. Video data (keypoints)
###############################################################
video_feature = scipy.io.loadmat('{0}/{1}.mat'.format(
self.cfg["video_feature_root"], sample_name))
# grab keypoints (discard foot keypoints)
video_feature = video_feature['lip_list'][0]
video_feature = video_feature[:, 0:self.cfg["num_lip_keypoints"], :]
# crop desired length from the video --- 300 frames for 12 seconds
n_frames = video_feature.shape[0]
frame_start = np.random.randint(
0, n_frames -
self.cfg["sample_duration"] * self.cfg["vid_framerate"] + 1)
frame_end = frame_start + self.cfg["sample_duration"] * self.cfg[
"vid_framerate"]
video_feature = video_feature[frame_start:frame_end, :, :]
# normalize keypoint positions
for kp in range(self.cfg["num_lip_keypoints"]):
scaler = MinMaxScaler(feature_range=(-1, 1))
video_feature[:,
kp, :] = scaler.fit_transform(video_feature[:,
kp, :])
#video_feature = video_feature.reshape(-1, self.cfg["num_lip_keypoints"]*2)
# video augmentation --- horizontal flipping
if self.mode == 'train':
if np.random.rand() < self.cfg["prob_horizontal_flip"]:
video_feature[:, :, 0] = video_feature[:, :, 0] * -1
###############################################################
# 2. Global Shift
###############################################################
if self.global_shift:
# shift the audio in feasible range
left_shift = min(frame_start, self.max_shift_frames)
right_shift = min(n_frames - frame_end, self.max_shift_frames)
frame_shift = np.random.randint(-left_shift, right_shift + 1)
audio_frame_start = frame_start + frame_shift
# modify the target array
target += 2 * frame_shift / (self.cfg["vid_framerate"] *
self.cfg["sample_duration"])
else:
audio_frame_start = frame_start
###############################################################
# 3. Audio data (Spectrogram)
###############################################################
waveform, audio_sample_rate = torchaudio.load('{0}/{1}.mp3'.format(
self.cfg["audio_feature_root"], sample_name))
# crop 12s of audio data
audio_start = int(audio_frame_start * self.cfg["mel_sr"] /
self.cfg["vid_framerate"])
audio_end = int(audio_start +
self.cfg["mel_sr"] * self.cfg["sample_duration"])
waveform = waveform[0, audio_start:audio_end]
# compute spectrogram
audio_feature = self.spec2db(self.calc_spec(waveform))
audio_feature = torch.squeeze(audio_feature, 0)
# normalize spectrogram
audio_feature = (audio_feature - torch.mean(audio_feature)) / (
torch.max(audio_feature) - torch.min(audio_feature))
# spectrogram augmentation
if self.mode == 'train':
# frequency masking
if np.random.rand() < self.cfg["prob_freqmask"]:
dur = np.random.randint(self.cfg["min_freqmask"],
self.cfg["max_freqmask"] + 1)
st = np.random.randint(0, audio_feature.shape[0] - dur + 1)
audio_feature[st:st + dur, :] = 0
# time masking
if np.random.rand() < self.cfg["prob_timemask"]:
dur = np.random.randint(self.cfg["min_timemask"],
self.cfg["max_timemask"] + 1)
st = np.random.randint(0, audio_feature.shape[1] - dur + 1)
audio_feature[:, st:st + dur] = 0
###############################################################
# 3. Global Stretch & Local Distortion
###############################################################
new_video_feature = np.zeros(
(self.frames, self.cfg["num_lip_keypoints"], 2))
new_target = np.zeros(self.frames)
# Random distortion
random_position = np.linspace(-1, 1, self.frames)
if self.local_distortion:
random_position = np.zeros(self.frames)
while np.max(
np.abs(random_position - np.linspace(-1, 1, self.frames))
) > self.max_distortion_ratio:
resample_len = int(self.frames *
self.cfg["random_resample_rate"])
random_position = np.random.rand(resample_len)
random_position = np.sort(
(random_position - np.min(random_position)) * 2 /
(np.max(random_position) - np.min(random_position)) - 1)
f = interpolate.interp1d(np.linspace(-1, 1, resample_len),
random_position,
kind='linear')
random_position = f(np.linspace(-1, 1, self.frames))
# Distorted & Stretched video feature
for k in range(self.frames):
orig_index = (random_position[k] + 1) / 2 * (
self.cfg["sample_duration"] * self.cfg["vid_framerate"] - 1)
lower = int(np.floor(orig_index))
upper = int(np.ceil(orig_index))
if lower == upper:
new_video_feature[k, :, :] = video_feature[lower, :, :]
new_target[k] = target[lower]
else:
new_video_feature[k, :, :] = video_feature[lower, :, :] * (
upper - orig_index) + video_feature[upper, :, :] * (
orig_index - lower)
new_target[k] = target[lower] * (
upper - orig_index) + target[upper] * (orig_index - lower)
video_feature = new_video_feature
target = new_target
# velocity
video_velo = np.zeros((self.frames, self.cfg["num_lip_keypoints"], 2))
video_velo[
1:, :, :] = video_feature[1:, :, :] - video_feature[:-1, :, :]
video_velo = video_velo / np.amax(np.absolute(video_velo))
# acceleration
video_acc = np.zeros((self.frames, self.cfg["num_lip_keypoints"], 2))
video_acc[1:, :, :] = video_velo[1:, :, :] - video_velo[:-1, :, :]
video_acc = video_acc / np.amax(np.absolute(video_acc))
# aggregate
video_agg = np.zeros(
(self.frames, self.cfg["num_lip_keypoints"], 2, 2))
video_agg[:, :, :, 0] = video_velo
video_agg[:, :, :, 1] = video_acc
return {"video_feature": torch.from_numpy(video_agg.astype(np.float32)), \
"audio_feature": audio_feature, \
"target": torch.from_numpy(target.astype(np.float32)), \
"sample_name": sample_name}
def extract_length(input_file):
wav, _ = torchaudio.load(input_file)
return wav.size(-1)
def torchaudio_load_file(file_path, normalization=True):
data, sr = torchaudio.load(str(file_path))
return data.float(), sr
def load_audio(self,audiopath):
self.waveform, self.sample_rate = torchaudio.load(audiopath)
return self.waveform,self.sample_rate
def download(self):
"""Download the yesno data if it doesn't exist in processed_folder already."""
from six.moves import urllib
import tarfile
if self._check_exists():
return
raw_abs_dir = os.path.join(self.root, self.raw_folder)
processed_abs_dir = os.path.join(self.root, self.processed_folder)
dset_abs_path = os.path.join(
self.root, self.raw_folder, self.dset_path)
# download files
try:
os.makedirs(os.path.join(self.root, self.raw_folder))
os.makedirs(os.path.join(self.root, self.processed_folder))
except OSError as e:
if e.errno == errno.EEXIST:
pass
else:
raise
url = self.url
print('Downloading ' + url)
filename = url.rpartition('/')[2]
file_path = os.path.join(self.root, self.raw_folder, filename)
if not os.path.isfile(file_path):
urllib.request.urlretrieve(url, file_path)
else:
print("Tar file already downloaded")
if not os.path.exists(dset_abs_path):
with tarfile.open(file_path) as zip_f:
zip_f.extractall(raw_abs_dir)
else:
print("Tar file already extracted")
if not self.dev_mode:
os.unlink(file_path)
# process and save as torch files
print('Processing...')
shutil.copyfile(
os.path.join(dset_abs_path, "README"),
os.path.join(processed_abs_dir, "YESNO_README")
)
audios = [x for x in os.listdir(dset_abs_path) if ".wav" in x]
print("Found {} audio files".format(len(audios)))
tensors = []
labels = []
lengths = []
for i, f in enumerate(audios):
full_path = os.path.join(dset_abs_path, f)
sig, sr = torchaudio.load(full_path)
tensors.append(sig)
lengths.append(sig.size(0))
labels.append(os.path.basename(f).split(".", 1)[0].split("_"))
# sort sigs/labels: longest -> shortest
tensors, labels = zip(*[(b, c) for (a, b, c) in sorted(
zip(lengths, tensors, labels), key=lambda x: x[0], reverse=True)])
self.max_len = tensors[0].size(0)
torch.save(
(tensors, labels),
os.path.join(
self.root,
self.processed_folder,
self.processed_file
)
)
if not self.dev_mode:
shutil.rmtree(raw_abs_dir, ignore_errors=True)
print('Done!')
def _load_wav(self, path):
wav, _ = torchaudio.load(path_join(self.data_dir, path))
wav = self.resampler(wav).squeeze(0)
return wav
def open(self, item, **kwargs):
if isinstance(item, (Path, PosixPath, str)):
sig, sr = torchaudio.load(item)
return AudioItem(sig, sr, path=Path(item))
if isinstance(item, (tuple, np.ndarray)):
return AudioItem(item)
# # To load audio data, you can use ``torchaudio.load``. # # This function accepts a path-like object or file-like object as input. # # The returned value is a tuple of waveform (``Tensor``) and sample rate # (``int``). # # By default, the resulting tensor object has ``dtype=torch.float32`` and # its value range is normalized within ``[-1.0, 1.0]``. # # For the list of supported format, please refer to `the torchaudio # documentation <https://pytorch.org/audio>`__. # waveform, sample_rate = torchaudio.load(SAMPLE_WAV_SPEECH_PATH) print_stats(waveform, sample_rate=sample_rate) plot_waveform(waveform, sample_rate) plot_specgram(waveform, sample_rate) play_audio(waveform, sample_rate) ###################################################################### # Loading from file-like object # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ # # ``torchaudio``\ ’s I/O functions now support file-like objects. This # allows for fetching and decoding audio data from locations # within and beyond the local file system. # The following examples illustrate this. #
def default_loader(path):
audio, sr = torchaudio.load(path)
return audio, sr
def create_gold_file(data_path, sample_rate):
"""
Create the following files:
gold_units.json : contains gold_dicts, a list of mappings of
{"sentence_id" : str,
"units" : a list of ints representing phoneme id for each feature frame,
"text" : a list of strs representing phoneme tokens for each feature frame}
abx_triplets.item : contains ABX triplets in the format
line 0 : whatever (not read)
line > 0: #file_ID onset offset #phone prev-phone next-phone speaker
onset : begining of the triplet (in s)
offset : end of the triplet (in s)
"""
wav_scp_file = os.path.join(data_path, "mscoco2k_wav.scp")
split_file = os.path.join(data_path, "mscoco2k_retrieval_split.txt")
select_idxs = [
idx for idx, is_test in enumerate(open(split_file, 'r'))
if int(is_test)
]
phone_info_dict = json.load(
open(os.path.join(data_path, "mscoco2k_phone_info.json"), "r"))
phone_to_index = {}
gold_dicts = []
triplets = ['#file_ID onset offset #phone prev-phone next-phone speaker']
# Extract audio file names as sentence ids
with open(wav_scp_file, 'r') as wav_scp_f:
filenames = [l.split()[-1] for idx, l in enumerate(wav_scp_f)]
# Extract utterance duration
durations = [
int(torchaudio.load(fn)[0].size(-1) * 1000 // (10 * sample_rate))
for fn in filenames
]
# Extract phone mapping
phone_path = os.path.join(data_path, "phone2id.json")
if os.path.exists(phone_path):
phone_to_index = json.load(open(phone_path, "r"))
else:
phones = set()
for idx, (_, phone_info) in enumerate(
sorted(phone_info_dict.items(),
key=lambda x: int(x[0].split("_")[-1]))):
for word_token in phone_info["data_ids"]:
for phone_token in word_token[2]:
token = phone_token[0]
phones.update([token])
phone_to_index = {x: i for i, x in enumerate(sorted(phones))}
phone_to_index[UNK] = len(phone_to_index)
json.dump(phone_to_index, open(phone_path, "w"), indent=2)
# Extract phone units
phone_to_word_counts = collections.defaultdict(dict)
global_idx = 0
for idx, (_, phone_info) in enumerate(
sorted(phone_info_dict.items(),
key=lambda x: int(x[0].split("_")[-1]))):
if not idx in select_idxs:
continue
begin_word = 0
for word_info, word_token in zip(phone_info["data_ids"],
phone_info["concepts"]):
dur_word = word_info[2][-1][2] - word_info[2][0][1]
end_word = begin_word + dur_word
nframes = int(dur_word // 10)
gold_dict = {
"sentence_id": filenames[idx],
"units": [-1] * nframes,
"phoneme_text": [UNK] * nframes,
"word_text": [word_token] * nframes,
"interval": [begin_word, end_word]
}
begin_phone = 0
prefix = filenames.split('/')[-1]
example_id = f"{prefix}_{global_idx}"
global_idx += 1
for phn_idx, phone_token in enumerate(word_info[2]):
if not word_token in phone_to_word_counts[phone_token[0]]:
phone_to_word_counts[phone_token[0]][word_token] = 1
else:
phone_to_word_counts[phone_token[0]][word_token] += 1
token, begin, end = phone_token[0], phone_token[
1], phone_token[2]
dur_phone = end - begin
begin_frame = int(begin_phone // 10)
end_frame = int((begin_phone + dur_phone) // 10)
if (begin_word + begin_phone +
dur_phone) // 10 > durations[idx]:
print(
'In {}: end frame exceeds duration of audio, {} > {}'.
format(filenames[idx],
(begin_word + begin_phone + dur_phone) // 10,
durations[idx]))
break
if phn_idx == 0:
prev_token = NULL
else:
prev_token = word_info[2][phn_idx - 1][0]
if phn_idx == len(word_info[2]) - 1:
next_token = NULL
else:
next_token = word_info[2][phn_idx + 1][0]
triplets.append(
f'{example_id} {begin_phone / 1000.0:.4f} {(begin_phone + dur_phone)/ 1000.0:.4f} {token} {prev_token} {next_token} 0'
)
for t in range(begin_frame, end_frame):
gold_dict["units"][t] = phone_to_index[token]
gold_dict["phoneme_text"][t] = token
begin_phone += dur_phone
if end_frame != nframes:
gold_dict['phoneme_text'] = gold_dict[
'phoneme_text'][:end_frame]
gold_dict['word_text'] = gold_dict['word_text'][:end_frame]
print('sentence_id, end_frame, nframes: ', filenames[idx],
end_frame, nframes)
gold_dicts.append(gold_dict)
begin_word += dur_word
with open(os.path.join(data_path, 'phone_token_top_10_words.txt'),
'w') as f:
f.write('Phone\tWord\tCounts\n')
for p in phone_to_word_counts:
for w in sorted(phone_to_word_counts[p],
key=lambda x: phone_to_word_counts[p][x],
reverse=True):
f.write('{}\t{}\t{}\n'.format(p, w,
phone_to_word_counts[p][w]))
with open(os.path.join(data_path, "gold_units.json"), "w") as gold_f:
json.dump(gold_dicts, gold_f, indent=2)
with open(os.path.join(data_path, "abx_triplets.item"), "w") as triplet_f:
f.write('\n'.join(triplets))
def test_batch_pitch(self): waveform, sample_rate = torchaudio.load(self.test_filepath) self._test_batch(F.detect_pitch_frequency, waveform, sample_rate)
import torch
import torchaudio
import matplotlib.pyplot as plt
torchaudio.set_audio_backend('soundfile')
waveform, sample_rate = torchaudio.load('data/Clover.flac')
print(f'Shape of waveform: {waveform.size()}')
print(f'Sample rate of waveform: {sample_rate}')
plt.figure()
plt.plot(waveform.t().numpy())
plt.show()
def test_jit_pitch(self): waveform, sample_rate = torchaudio.load(self.test_filepath) _test_torchscript_functional(F.detect_pitch_frequency, waveform, sample_rate)
def test_save(self):
# load signal
x, sr = torchaudio.load(self.test_filepath)
# check save
new_filepath = os.path.join(self.test_dirpath, "test.wav")
torchaudio.save(new_filepath, x, sr)
self.assertTrue(os.path.isfile(new_filepath))
os.unlink(new_filepath)
# check automatic normalization
x /= 1 << 31
torchaudio.save(new_filepath, x, sr)
self.assertTrue(os.path.isfile(new_filepath))
os.unlink(new_filepath)
# test save 1d tensor
x = x[:, 0] # get mono signal
x.squeeze_() # remove channel dim
torchaudio.save(new_filepath, x, sr)
self.assertTrue(os.path.isfile(new_filepath))
os.unlink(new_filepath)
# don't allow invalid sizes as inputs
with self.assertRaises(ValueError):
x.unsqueeze_(0) # N x L not L x N
torchaudio.save(new_filepath, x, sr)
with self.assertRaises(ValueError):
x.squeeze_()
x.unsqueeze_(1)
x.unsqueeze_(0) # 1 x L x 1
torchaudio.save(new_filepath, x, sr)
# automatically convert sr from floating point to int
x.squeeze_(0)
torchaudio.save(new_filepath, x, float(sr))
self.assertTrue(os.path.isfile(new_filepath))
os.unlink(new_filepath)
# don't allow uneven integers
with self.assertRaises(TypeError):
torchaudio.save(new_filepath, x, float(sr) + 0.5)
self.assertTrue(os.path.isfile(new_filepath))
os.unlink(new_filepath)
# don't save to folders that don't exist
with self.assertRaises(OSError):
new_filepath = os.path.join(self.test_dirpath, "no-path",
"test.wav")
torchaudio.save(new_filepath, x, sr)
# save created file
sinewave_filepath = os.path.join(self.test_dirpath, "assets",
"sinewave.wav")
sr = 16000
freq = 440
volume = 0.3
y = (torch.cos(2 * math.pi * torch.arange(0, 4 * sr).float() * freq /
sr))
y.unsqueeze_(1)
# y is between -1 and 1, so must scale
y = (y * volume * 2**31).long()
torchaudio.save(sinewave_filepath, y, sr)
self.assertTrue(os.path.isfile(sinewave_filepath))
# test precision
new_filepath = os.path.join(self.test_dirpath, "test.wav")
_, _, _, bp = torchaudio.info(sinewave_filepath)
torchaudio.save(new_filepath, y, sr, precision=16)
_, _, _, bp16 = torchaudio.info(new_filepath)
self.assertEqual(bp, 32)
self.assertEqual(bp16, 16)
os.unlink(new_filepath)
class TestFunctional(unittest.TestCase):
data_sizes = [(2, 20), (3, 15), (4, 10)]
number_of_trials = 100
specgram = torch.tensor([1., 2., 3., 4.])
test_dirpath, test_dir = common_utils.create_temp_assets_dir()
test_filepath = os.path.join(test_dirpath, 'assets',
'steam-train-whistle-daniel_simon.mp3')
waveform_train, sr_train = torchaudio.load(test_filepath)
def test_torchscript_spectrogram(self):
tensor = torch.rand((1, 1000))
n_fft = 400
ws = 400
hop = 200
pad = 0
window = torch.hann_window(ws)
power = 2
normalize = False
_test_torchscript_functional(
F.spectrogram, tensor, pad, window, n_fft, hop, ws, power, normalize
)
def test_torchscript_griffinlim(self):
tensor = torch.rand((1, 201, 6))
n_fft = 400
ws = 400
hop = 200
window = torch.hann_window(ws)
power = 2
normalize = False
momentum = 0.99
n_iter = 32
length = 1000
init = 0
_test_torchscript_functional(
F.griffinlim, tensor, window, n_fft, hop, ws, power, normalize, n_iter, momentum, length, 0
)
def test_batch_griffinlim(self):
torch.random.manual_seed(42)
tensor = torch.rand((1, 201, 6))
n_fft = 400
ws = 400
hop = 200
window = torch.hann_window(ws)
power = 2
normalize = False
momentum = 0.99
n_iter = 32
length = 1000
self._test_batch(
F.griffinlim, tensor, window, n_fft, hop, ws, power, normalize, n_iter, momentum, length, 0, atol=5e-5
)
def _test_compute_deltas(self, specgram, expected, win_length=3, atol=1e-6, rtol=1e-8):
computed = F.compute_deltas(specgram, win_length=win_length)
self.assertTrue(computed.shape == expected.shape, (computed.shape, expected.shape))
torch.testing.assert_allclose(computed, expected, atol=atol, rtol=rtol)
def test_compute_deltas_onechannel(self):
specgram = self.specgram.unsqueeze(0).unsqueeze(0)
expected = torch.tensor([[[0.5, 1.0, 1.0, 0.5]]])
self._test_compute_deltas(specgram, expected)
def test_compute_deltas_twochannel(self):
specgram = self.specgram.repeat(1, 2, 1)
expected = torch.tensor([[[0.5, 1.0, 1.0, 0.5],
[0.5, 1.0, 1.0, 0.5]]])
self._test_compute_deltas(specgram, expected)
def test_compute_deltas_randn(self):
channel = 13
n_mfcc = channel * 3
time = 1021
win_length = 2 * 7 + 1
specgram = torch.randn(channel, n_mfcc, time)
computed = F.compute_deltas(specgram, win_length=win_length)
self.assertTrue(computed.shape == specgram.shape, (computed.shape, specgram.shape))
_test_torchscript_functional(F.compute_deltas, specgram, win_length=win_length)
def test_batch_pitch(self):
waveform, sample_rate = torchaudio.load(self.test_filepath)
self._test_batch(F.detect_pitch_frequency, waveform, sample_rate)
def test_jit_pitch(self):
waveform, sample_rate = torchaudio.load(self.test_filepath)
_test_torchscript_functional(F.detect_pitch_frequency, waveform, sample_rate)
def _compare_estimate(self, sound, estimate, atol=1e-6, rtol=1e-8):
# trim sound for case when constructed signal is shorter than original
sound = sound[..., :estimate.size(-1)]
self.assertTrue(sound.shape == estimate.shape, (sound.shape, estimate.shape))
self.assertTrue(torch.allclose(sound, estimate, atol=atol, rtol=rtol))
def _test_istft_is_inverse_of_stft(self, kwargs):
# generates a random sound signal for each tril and then does the stft/istft
# operation to check whether we can reconstruct signal
for data_size in self.data_sizes:
for i in range(self.number_of_trials):
sound = common_utils.random_float_tensor(i, data_size)
stft = torch.stft(sound, **kwargs)
estimate = torchaudio.functional.istft(stft, length=sound.size(1), **kwargs)
self._compare_estimate(sound, estimate)
def test_istft_is_inverse_of_stft1(self):
# hann_window, centered, normalized, onesided
kwargs1 = {
'n_fft': 12,
'hop_length':4,
'win_length':12,
'window': torch.hann_window(12),
'center': True,
'pad_mode': 'reflect',
'normalized':True,
'onesided': True,
}
self._test_istft_is_inverse_of_stft(kwargs1)
def test_istft_is_inverse_of_stft2(self):
# hann_window, centered, not normalized, not onesided
kwargs2 = {
'n_fft': 12,
'hop_length':2,
'win_length':8,
'window': torch.hann_window(8),
'center': True,
'pad_mode': 'reflect',
'normalized':False,
'onesided': False,
}
self._test_istft_is_inverse_of_stft(kwargs2)
def test_istft_is_inverse_of_stft3(self):
# hamming_window, centered, normalized, not onesided
kwargs3 = {
'n_fft': 15,
'hop_length':3,
'win_length':11,
'window': torch.hamming_window(11),
'center': True,
'pad_mode': 'constant',
'normalized':True,
'onesided': False,
}
self._test_istft_is_inverse_of_stft(kwargs3)
def test_istft_is_inverse_of_stft4(self):
# hamming_window, not centered, not normalized, onesided
# window same size as n_fft
kwargs4 = {
'n_fft': 5,
'hop_length':2,
'win_length':5,
'window': torch.hamming_window(5),
'center': False,
'pad_mode': 'constant',
'normalized':False,
'onesided': True,
}
self._test_istft_is_inverse_of_stft(kwargs4)
def test_istft_is_inverse_of_stft5(self):
# hamming_window, not centered, not normalized, not onesided
# window same size as n_fft
kwargs5 = {
'n_fft': 3,
'hop_length':2,
'win_length':3,
'window': torch.hamming_window(3),
'center': False,
'pad_mode': 'reflect',
'normalized':False,
'onesided': False,
}
self._test_istft_is_inverse_of_stft(kwargs5)
def test_istft_of_ones(self):
# stft = torch.stft(torch.ones(4), 4)
stft = torch.tensor([
[[4., 0.], [4., 0.], [4., 0.], [4., 0.], [4., 0.]],
[[0., 0.], [0., 0.], [0., 0.], [0., 0.], [0., 0.]],
[[0., 0.], [0., 0.], [0., 0.], [0., 0.], [0., 0.]]
])
estimate = torchaudio.functional.istft(stft, n_fft=4, length=4)
self._compare_estimate(torch.ones(4), estimate)
def test_istft_of_zeros(self):
# stft = torch.stft(torch.zeros(4), 4)
stft = torch.zeros((3, 5, 2))
estimate = torchaudio.functional.istft(stft, n_fft=4, length=4)
self._compare_estimate(torch.zeros(4), estimate)
def test_istft_requires_overlap_windows(self):
# the window is size 1 but it hops 20 so there is a gap which throw an error
stft = torch.zeros((3, 5, 2))
self.assertRaises(AssertionError, torchaudio.functional.istft, stft, n_fft=4,
hop_length=20, win_length=1, window=torch.ones(1))
def test_istft_requires_nola(self):
stft = torch.zeros((3, 5, 2))
kwargs_ok = {
'n_fft': 4,
'win_length':4,
'window': torch.ones(4),
}
kwargs_not_ok = {
'n_fft': 4,
'win_length':4,
'window': torch.zeros(4),
}
# A window of ones meets NOLA but a window of zeros does not. This should
# throw an error.
torchaudio.functional.istft(stft, **kwargs_ok)
self.assertRaises(AssertionError, torchaudio.functional.istft, stft, **kwargs_not_ok)
def test_istft_requires_non_empty(self):
self.assertRaises(AssertionError, torchaudio.functional.istft, torch.zeros((3, 0, 2)), 2)
self.assertRaises(AssertionError, torchaudio.functional.istft, torch.zeros((0, 3, 2)), 2)
def _test_istft_of_sine(self, amplitude, L, n):
# stft of amplitude*sin(2*pi/L*n*x) with the hop length and window size equaling L
x = torch.arange(2 * L + 1, dtype=torch.get_default_dtype())
sound = amplitude * torch.sin(2 * math.pi / L * x * n)
# stft = torch.stft(sound, L, hop_length=L, win_length=L,
# window=torch.ones(L), center=False, normalized=False)
stft = torch.zeros((L // 2 + 1, 2, 2))
stft_largest_val = (amplitude * L) / 2.0
if n < stft.size(0):
stft[n, :, 1] = -stft_largest_val
if 0 <= L - n < stft.size(0):
# symmetric about L // 2
stft[L - n, :, 1] = stft_largest_val
estimate = torchaudio.functional.istft(stft, L, hop_length=L, win_length=L,
window=torch.ones(L), center=False, normalized=False)
# There is a larger error due to the scaling of amplitude
self._compare_estimate(sound, estimate, atol=1e-3)
def test_istft_of_sine(self):
self._test_istft_of_sine(amplitude=123, L=5, n=1)
self._test_istft_of_sine(amplitude=150, L=5, n=2)
self._test_istft_of_sine(amplitude=111, L=5, n=3)
self._test_istft_of_sine(amplitude=160, L=7, n=4)
self._test_istft_of_sine(amplitude=145, L=8, n=5)
self._test_istft_of_sine(amplitude=80, L=9, n=6)
self._test_istft_of_sine(amplitude=99, L=10, n=7)
def _test_linearity_of_istft(self, data_size, kwargs, atol=1e-6, rtol=1e-8):
for i in range(self.number_of_trials):
tensor1 = common_utils.random_float_tensor(i, data_size)
tensor2 = common_utils.random_float_tensor(i * 2, data_size)
a, b = torch.rand(2)
istft1 = torchaudio.functional.istft(tensor1, **kwargs)
istft2 = torchaudio.functional.istft(tensor2, **kwargs)
istft = a * istft1 + b * istft2
estimate = torchaudio.functional.istft(a * tensor1 + b * tensor2, **kwargs)
self._compare_estimate(istft, estimate, atol, rtol)
def test_linearity_of_istft1(self):
# hann_window, centered, normalized, onesided
kwargs1 = {
'n_fft': 12,
'window': torch.hann_window(12),
'center': True,
'pad_mode': 'reflect',
'normalized':True,
'onesided': True,
}
data_size = (2, 7, 7, 2)
self._test_linearity_of_istft(data_size, kwargs1)
def test_linearity_of_istft2(self):
# hann_window, centered, not normalized, not onesided
kwargs2 = {
'n_fft': 12,
'window': torch.hann_window(12),
'center': True,
'pad_mode': 'reflect',
'normalized':False,
'onesided': False,
}
data_size = (2, 12, 7, 2)
self._test_linearity_of_istft(data_size, kwargs2)
def test_linearity_of_istft3(self):
# hamming_window, centered, normalized, not onesided
kwargs3 = {
'n_fft': 12,
'window': torch.hamming_window(12),
'center': True,
'pad_mode': 'constant',
'normalized':True,
'onesided': False,
}
data_size = (2, 12, 7, 2)
self._test_linearity_of_istft(data_size, kwargs3)
def test_linearity_of_istft4(self):
# hamming_window, not centered, not normalized, onesided
kwargs4 = {
'n_fft': 12,
'window': torch.hamming_window(12),
'center': False,
'pad_mode': 'constant',
'normalized':False,
'onesided': True,
}
data_size = (2, 7, 3, 2)
self._test_linearity_of_istft(data_size, kwargs4, atol=1e-5, rtol=1e-8)
def test_batch_istft(self):
stft = torch.tensor([
[[4., 0.], [4., 0.], [4., 0.], [4., 0.], [4., 0.]],
[[0., 0.], [0., 0.], [0., 0.], [0., 0.], [0., 0.]],
[[0., 0.], [0., 0.], [0., 0.], [0., 0.], [0., 0.]]
])
self._test_batch(F.istft, stft, n_fft=4, length=4)
def _test_create_fb(self, n_mels=40, sample_rate=22050, n_fft=2048, fmin=0.0, fmax=8000.0):
librosa_fb = librosa.filters.mel(sr=sample_rate,
n_fft=n_fft,
n_mels=n_mels,
fmax=fmax,
fmin=fmin,
htk=True,
norm=None)
fb = F.create_fb_matrix(sample_rate=sample_rate,
n_mels=n_mels,
f_max=fmax,
f_min=fmin,
n_freqs=(n_fft // 2 + 1))
for i_mel_bank in range(n_mels):
assert torch.allclose(fb[:, i_mel_bank], torch.tensor(librosa_fb[i_mel_bank]), atol=1e-4)
def test_create_fb(self):
self._test_create_fb()
self._test_create_fb(n_mels=128, sample_rate=44100)
self._test_create_fb(n_mels=128, fmin=2000.0, fmax=5000.0)
self._test_create_fb(n_mels=56, fmin=100.0, fmax=9000.0)
self._test_create_fb(n_mels=56, fmin=800.0, fmax=900.0)
self._test_create_fb(n_mels=56, fmin=1900.0, fmax=900.0)
self._test_create_fb(n_mels=10, fmin=1900.0, fmax=900.0)
def test_gain(self):
waveform_gain = F.gain(self.waveform_train, 3)
self.assertTrue(waveform_gain.abs().max().item(), 1.)
E = torchaudio.sox_effects.SoxEffectsChain()
E.set_input_file(self.test_filepath)
E.append_effect_to_chain("gain", [3])
sox_gain_waveform = E.sox_build_flow_effects()[0]
self.assertTrue(torch.allclose(waveform_gain, sox_gain_waveform, atol=1e-04))
def test_dither(self):
waveform_dithered = F.dither(self.waveform_train)
waveform_dithered_noiseshaped = F.dither(self.waveform_train, noise_shaping=True)
E = torchaudio.sox_effects.SoxEffectsChain()
E.set_input_file(self.test_filepath)
E.append_effect_to_chain("dither", [])
sox_dither_waveform = E.sox_build_flow_effects()[0]
self.assertTrue(torch.allclose(waveform_dithered, sox_dither_waveform, atol=1e-04))
E.clear_chain()
E.append_effect_to_chain("dither", ["-s"])
sox_dither_waveform_ns = E.sox_build_flow_effects()[0]
self.assertTrue(torch.allclose(waveform_dithered_noiseshaped, sox_dither_waveform_ns, atol=1e-02))
def test_vctk_transform_pipeline(self):
test_filepath_vctk = os.path.join(self.test_dirpath, "assets/VCTK-Corpus/wav48/p224/", "p224_002.wav")
wf_vctk, sr_vctk = torchaudio.load(test_filepath_vctk)
# rate
sample = T.Resample(sr_vctk, 16000, resampling_method='sinc_interpolation')
wf_vctk = sample(wf_vctk)
# dither
wf_vctk = F.dither(wf_vctk, noise_shaping=True)
E = torchaudio.sox_effects.SoxEffectsChain()
E.set_input_file(test_filepath_vctk)
E.append_effect_to_chain("gain", ["-h"])
E.append_effect_to_chain("channels", [1])
E.append_effect_to_chain("rate", [16000])
E.append_effect_to_chain("gain", ["-rh"])
E.append_effect_to_chain("dither", ["-s"])
wf_vctk_sox = E.sox_build_flow_effects()[0]
self.assertTrue(torch.allclose(wf_vctk, wf_vctk_sox, rtol=1e-03, atol=1e-03))
def test_pitch(self):
test_dirpath, test_dir = common_utils.create_temp_assets_dir()
test_filepath_100 = os.path.join(test_dirpath, 'assets', "100Hz_44100Hz_16bit_05sec.wav")
test_filepath_440 = os.path.join(test_dirpath, 'assets', "440Hz_44100Hz_16bit_05sec.wav")
# Files from https://www.mediacollege.com/audio/tone/download/
tests = [
(test_filepath_100, 100),
(test_filepath_440, 440),
]
for filename, freq_ref in tests:
waveform, sample_rate = torchaudio.load(filename)
freq = torchaudio.functional.detect_pitch_frequency(waveform, sample_rate)
threshold = 1
s = ((freq - freq_ref).abs() > threshold).sum()
self.assertFalse(s)
# Convert to stereo and batch for testing purposes
self._test_batch(F.detect_pitch_frequency, waveform, sample_rate)
def _test_batch_shape(self, functional, tensor, *args, **kwargs):
kwargs_compare = {}
if 'atol' in kwargs:
atol = kwargs['atol']
del kwargs['atol']
kwargs_compare['atol'] = atol
if 'rtol' in kwargs:
rtol = kwargs['rtol']
del kwargs['rtol']
kwargs_compare['rtol'] = rtol
# Single then transform then batch
torch.random.manual_seed(42)
expected = functional(tensor.clone(), *args, **kwargs)
expected = expected.unsqueeze(0).unsqueeze(0)
# 1-Batch then transform
tensors = tensor.unsqueeze(0).unsqueeze(0)
torch.random.manual_seed(42)
computed = functional(tensors.clone(), *args, **kwargs)
self._compare_estimate(computed, expected, **kwargs_compare)
return tensors, expected
def _test_batch(self, functional, tensor, *args, **kwargs):
tensors, expected = self._test_batch_shape(functional, tensor, *args, **kwargs)
kwargs_compare = {}
if 'atol' in kwargs:
atol = kwargs['atol']
del kwargs['atol']
kwargs_compare['atol'] = atol
if 'rtol' in kwargs:
rtol = kwargs['rtol']
del kwargs['rtol']
kwargs_compare['rtol'] = rtol
# 3-Batch then transform
ind = [3] + [1] * (int(tensors.dim()) - 1)
tensors = tensor.repeat(*ind)
ind = [3] + [1] * (int(expected.dim()) - 1)
expected = expected.repeat(*ind)
torch.random.manual_seed(42)
computed = functional(tensors.clone(), *args, **kwargs)
def test_torchscript_create_fb_matrix(self):
n_stft = 100
f_min = 0.0
f_max = 20.0
n_mels = 10
sample_rate = 16000
_test_torchscript_functional(F.create_fb_matrix, n_stft, f_min, f_max, n_mels, sample_rate)
def test_torchscript_amplitude_to_DB(self):
spec = torch.rand((6, 201))
multiplier = 10.0
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
_test_torchscript_functional(F.amplitude_to_DB, spec, multiplier, amin, db_multiplier, top_db)
def test_torchscript_DB_to_amplitude(self):
x = torch.rand((1, 100))
ref = 1.
power = 1.
_test_torchscript_functional(F.DB_to_amplitude, x, ref, power)
def test_DB_to_amplitude(self):
# Make some noise
x = torch.rand(1000)
spectrogram = torchaudio.transforms.Spectrogram()
spec = spectrogram(x)
amin = 1e-10
ref = 1.0
db_multiplier = math.log10(max(amin, ref))
# Waveform amplitude -> DB -> amplitude
multiplier = 20.
power = 0.5
db = F.amplitude_to_DB(torch.abs(x), multiplier, amin, db_multiplier, top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertTrue(torch.allclose(torch.abs(x), x2, atol=5e-5))
# Spectrogram amplitude -> DB -> amplitude
db = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertTrue(torch.allclose(spec, x2, atol=5e-5))
# Waveform power -> DB -> power
multiplier = 10.
power = 1.
db = F.amplitude_to_DB(x, multiplier, amin, db_multiplier, top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertTrue(torch.allclose(torch.abs(x), x2, atol=5e-5))
# Spectrogram power -> DB -> power
db = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertTrue(torch.allclose(spec, x2, atol=5e-5))
def test_torchscript_create_dct(self):
n_mfcc = 40
n_mels = 128
norm = "ortho"
_test_torchscript_functional(F.create_dct, n_mfcc, n_mels, norm)
def test_torchscript_mu_law_encoding(self):
tensor = torch.rand((1, 10))
qc = 256
_test_torchscript_functional(F.mu_law_encoding, tensor, qc)
def test_torchscript_mu_law_decoding(self):
tensor = torch.rand((1, 10))
qc = 256
_test_torchscript_functional(F.mu_law_decoding, tensor, qc)
def test_torchscript_complex_norm(self):
complex_tensor = torch.randn(1, 2, 1025, 400, 2)
power = 2
_test_torchscript_functional(F.complex_norm, complex_tensor, power)
def test_mask_along_axis(self):
specgram = torch.randn(2, 1025, 400)
mask_param = 100
mask_value = 30.
axis = 2
_test_torchscript_functional(F.mask_along_axis, specgram, mask_param, mask_value, axis)
def test_mask_along_axis_iid(self):
specgrams = torch.randn(4, 2, 1025, 400)
mask_param = 100
mask_value = 30.
axis = 2
_test_torchscript_functional(F.mask_along_axis_iid, specgrams, mask_param, mask_value, axis)
def test_torchscript_gain(self):
tensor = torch.rand((1, 1000))
gainDB = 2.0
_test_torchscript_functional(F.gain, tensor, gainDB)
def test_torchscript_dither(self):
tensor = torch.rand((2, 1000))
_test_torchscript_functional_shape(F.dither, tensor)
_test_torchscript_functional_shape(F.dither, tensor, "RPDF")
_test_torchscript_functional_shape(F.dither, tensor, "GPDF")
def test_Vad(self):
filepath = common_utils.get_asset_path("vad-go-mono-32000.wav")
waveform, sample_rate = torchaudio.load(filepath)
self._assert_consistency(T.Vad(sample_rate=sample_rate), waveform)
def load_wav(self, file):
waveform, sample_rate = torchaudio.load(self.wav_folder + file)
if sample_rate != self.sr:
raise (AssertionError)
return waveform, sample_rate
def load_audio(path):
waveform, sample_rate = torchaudio.load(path)
return waveform
f_info[0] : speaker ID
f_info[1] : session number
f_info[2] : sentence number
f_info[3] : extension
"""
print("===extract features to .npy===")
for gender in ["male", "female"]:
SUBDIR_PATH = DIR_PATH / gender
for f in SUBDIR_PATH.rglob('*.wav'):
f_info = re.split('[_.]', f.name)
if int(f_info[2]) > 30:
continue
waveform, sample_rate = torchaudio.load(SUBDIR_PATH / f_info[0] /
f.name,
normalization=True)
feature = mfcc(waveform,
sample_rate,
winfunc=np.hamming,
numcep=hyper_params["numcep"],
nfilt=26,
nfft=512,
preemph=0.97)
speaker = str(f_info[0])
if speaker in train_speaker_list:
f_path = FEATURE_PATH / "train" / f.name[:-4]
elif speaker in val_speaker_list:
from pase.models.frontend import wf_builder
pase = wf_builder('cfg/frontend/PASE+.cfg').eval()
pase.load_pretrained('FE_e199.ckpt', load_last=True, verbose=True)
pase.cuda()
import sys
wav_path = sys.argv[1]
out_path = sys.argv[2]
# Now we can forward waveforms as Torch tensors
import torch
import torchaudio
torchaudio.set_audio_backend('sox')
x, sr = torchaudio.load(wav_path)
x = x.view(-1).cuda()
x = x.view(1, 1, -1)
with torch.no_grad():
pase.eval()
# y size will be (1, 256, 625), which are 625 frames of 256 dims each
y = pase(x)[0].transpose(0, 1)
torch.save(y.detach().cpu(), out_path)
#build dataset
dataset_path = "D:/songs_headphones/stereo-cut-16k"
n_files = 0
fs = 0
pad = nn.ConstantPad1d((0, 1), 0)
start_time = time.time()
for subdir, dirs, files in os.walk(dataset_path):
for file in files:
filepath = subdir + os.sep + file
if filepath.endswith(".wav"):
if n_files == 0:
dataset, fs = torchaudio.load(filepath)
dataset = pad(dataset)
dataset = dataset / torch.max(dataset)
elif n_files == 1:
new_file, fs = torchaudio.load(filepath)
new_file = pad(new_file)
new_file = new_file / torch.max(new_file)
dataset = torch.stack((dataset, new_file))
else:
new_file, fs = torchaudio.load(filepath)
new_file = pad(new_file)
new_file = new_file / torch.max(new_file)
dataset = torch.cat((dataset, new_file.unsqueeze(0)))
padding=False,
num_padding=0)
wav_file = "/d1/jbaik/ics-asr/temp/conan1-8k.wav"
audio = transformer(wav_file)
# test Spectrogram
elif test == 2:
import matplotlib
matplotlib.use('TkAgg')
matplotlib.interactive(True)
import matplotlib.pyplot as plt
nperseg = int(p.SAMPLE_RATE * p.WINDOW_SIZE)
noverlap = int(p.SAMPLE_RATE * (p.WINDOW_SIZE - p.WINDOW_SHIFT))
wav_file = Path("../data/aspire/000/fe_03_00047-A-025005-025135.wav")
audio, _ = torchaudio.load(wav_file)
# pyplot specgram
audio = torch.squeeze(audio)
fig = plt.figure(0)
plt.specgram(audio,
Fs=p.SAMPLE_RATE,
NFFT=p.NFFT,
noverlap=noverlap,
cmap='plasma')
# implemented transformer - scipy stft
transformer = Spectrogram(sample_rate=p.SAMPLE_RATE,
window_stride=p.WINDOW_SHIFT,
window_size=p.WINDOW_SIZE,
nfft=p.NFFT)
def read_audio_file(path, src_dir, side, sample_rate, window_size,
window_stride, window, normalize_audio,
truncate=None):
"""
Args:
path (str): location of a src file containing audio paths.
src_dir (str): location of source audio files.
side (str): 'src' or 'tgt'.
sample_rate (int): sample_rate.
window_size (float) : window size for spectrogram in seconds.
window_stride (float): window stride for spectrogram in seconds.
window (str): window type for spectrogram generation.
normalize_audio (bool): subtract spectrogram by mean and divide
by std or not.
truncate (int): maximum audio length (0 or None for unlimited).
Yields:
a dictionary containing audio data for each line.
"""
assert (src_dir is not None) and os.path.exists(src_dir),\
"src_dir must be a valid directory if data_type is audio"
global torchaudio, librosa, np
import torchaudio
import librosa
import numpy as np
with codecs.open(path, "r", "utf-8") as corpus_file:
index = 0
for line in corpus_file:
audio_path = os.path.join(src_dir, line.strip())
if not os.path.exists(audio_path):
audio_path = line
assert os.path.exists(audio_path), \
'audio path %s not found' % (line.strip())
sound, sample_rate = torchaudio.load(audio_path)
if truncate and truncate > 0:
if sound.size(0) > truncate:
continue
assert sample_rate == sample_rate, \
'Sample rate of %s != -sample_rate (%d vs %d)' \
% (audio_path, sample_rate, sample_rate)
sound = sound.numpy()
if len(sound.shape) > 1:
if sound.shape[1] == 1:
sound = sound.squeeze()
else:
sound = sound.mean(axis=1) # average multiple channels
n_fft = int(sample_rate * window_size)
win_length = n_fft
hop_length = int(sample_rate * window_stride)
# STFT
d = librosa.stft(sound, n_fft=n_fft, hop_length=hop_length,
win_length=win_length, window=window)
spect, _ = librosa.magphase(d)
spect = np.log1p(spect)
spect = torch.FloatTensor(spect)
if normalize_audio:
mean = spect.mean()
std = spect.std()
spect.add_(-mean)
spect.div_(std)
example_dict = {side: spect,
side + '_path': line.strip(),
'indices': index}
index += 1
yield example_dict