def get_mel(self, filename):
if not self.load_mel_from_disk:
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.stft.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, self.stft.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm,
requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
else:
melspec = torch.from_numpy(np.load(filename))
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
return melspec
def __getitem__(self, index):
filename = self.audio_files[index]
filename = os.path.join('dataset', filename)
audio, sampling_rate = utils.load_wav_to_torch(filename)
if sampling_rate != self.sampling_rate:
print(filename)
raise ValueError("Sampling rate doesn't math")
if audio.size(0) >= self.segment_length:
max_audio_start = audio.size(0) - self.segment_length
audio_start = random.randint(0, max_audio_start)
audio = audio[audio_start:audio_start + self.segment_length]
else:
audio = F.pad(audio, (0, self.segment_length - audio.size(0)),
'constant').data
audio = utils.mu_law_encode(audio / utils.MAX_WAV_VALUE,
self.mu_quantization)
return audio
def get_mel(self, filename, stft):
try:
melspec = torch.from_numpy(np.load(filename[:-4] + '_' + str(stft.n_mel_channels) + '.npy'))
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
except:
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != stft.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, stft.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
with open(filename[:-4] + '_' + str(stft.n_mel_channels) + '.npy', 'wb+') as f:
np.save(f, melspec.numpy())
return melspec
def get_mel(self, filename):
if not self.load_mel_from_disk:
audio, sampling_rate, max_value = load_wav_to_torch(filename)
if self.audio_offset: # used for extreme GTA'ing
audio = audio[self.audio_offset:]
self.max_wav_value = max(max_value, audio.max().item(), -audio.min().item()) # I'm not sure how, but sometimes the magnitude of audio exceeds the max of the datatype used before casting.
if sampling_rate != self.stft.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, self.stft.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
else:
melspec = torch.from_numpy(np.load(filename, allow_pickle=True)).float()
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
return melspec
def get_audio(self, filename):
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, self.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
spec_filename = filename.replace(".wav", ".spec.pt")
if os.path.exists(spec_filename):
spec = torch.load(spec_filename)
else:
spec = spectrogram_torch(audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename)
return spec, audio_norm
def get_mel_and_f0(self, filepath):
audio, sampling_rate = load_wav_to_torch(filepath)
audio_norm = audio / self.max_wav_value
# if sampling_rate != self.stft.sampling_rate:
# raise ValueError("{} SR doesn't match target {} SR".format(
# sampling_rate, self.stft.sampling_rate))
# audio_norm = audio_norm.unsqueeze(0)
# melspec = self.stft.mel_spectrogram(audio_norm)
# melspec = torch.squeeze(melspec, 0)
melspec = linearspectrogram_torch(audio_norm) # 用aukit的频谱生成方案
f0 = self.get_f0(audio.cpu().numpy(), sampling_rate,
self.filter_length, self.hop_length, self.f0_min,
self.f0_max, self.harm_thresh)
f0 = torch.from_numpy(f0)[None]
# f0 = f0[:, :melspec.size(1)]
# 用零向量替换F0
# f0 = torch.zeros(1, melspec.shape[1], dtype=torch.float)
return melspec, f0
def get_mel_and_f0(filepath, filter_length, hop_length, win_length, n_mel_channels, sampling_rate, mel_fmin, mel_fmax, f0_min, f0_max, harm_thresh):
stft = layers.TacotronSTFT(filter_length, hop_length, win_length,
n_mel_channels, sampling_rate, mel_fmin,
mel_fmax)
audio, sampling_rate = load_wav_to_torch(filepath)
if sampling_rate != stft.sampling_rate:
raise ValueError("{} SR doesn't match target {} SR".format(
sampling_rate, stft.sampling_rate))
audio_norm = audio
# I changed them to float32 during preprocessing so this normalization is unnecessary.
audio_norm = audio_norm.unsqueeze(0)
melspec = stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
f0 = get_f0(audio.cpu().numpy(), sampling_rate,
filter_length, hop_length, f0_min,
f0_max, harm_thresh)
f0 = torch.from_numpy(f0)[None]
f0 = f0[:, :melspec.size(1)]
return melspec, f0
def __getitem__(self, index):
# Read audio
filename = self.audio_files[index]
audio, sampling_rate = utils.load_wav_to_torch(filename)
if sampling_rate != self.sampling_rate:
raise ValueError("{} SR doesn't match target {} SR".format(
sampling_rate, self.sampling_rate))
# Take segment
if audio.size(0) >= self.segment_length:
max_audio_start = audio.size(0) - self.segment_length
audio_start = random.randint(0, max_audio_start)
audio = audio[audio_start:audio_start + self.segment_length]
else:
audio = torch.nn.functional.pad(
audio, (0, self.segment_length - audio.size(0)),
'constant').data
mel = self.get_mel(audio)
audio = utils.mu_law_encode(audio / utils.MAX_WAV_VALUE,
self.mu_quantization)
return (mel, audio)
def get_mel(self, fb):
if self.load_mel_from_disk:
cur_mel_path = os.path.join(self.mel_path, fb + '.npy')
melspec = np.load(cur_mel_path)
mean, std = np.load(self.MelStd_mel)
melspec = (melspec - mean) / std
melspec = np.transpose(melspec)
melspec = torch.from_numpy(melspec)
else:
cur_audio_path = os.path.join(self.audio_path, fb + '.wav')
audio = load_wav_to_torch(cur_audio_path, self.sampling_rate)
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm,
requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0) # [mel_bin, T]
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
return melspec
def prepare_mel_npy(hparams, audiopath_and_text):
audiopath_and_texts = load_filepaths_and_text(audiopath_and_text)
for i in range(len(audiopath_and_texts)):
audiopath, text, speaker_id = audiopath_and_texts[i][0], audiopath_and_texts[i][1], audiopath_and_texts[i][2]
audio, sampling_rate = load_wav_to_torch(audiopath)
audio_norm = audio / hparams.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = layers.TacotronSTFT(hparams).mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
out_dir = audiopath[:11]
file_name = audiopath[12:-4]
file = os.path.join(out_dir, file_name)
np.save(file, melspec)
print("{} / {}".format(i,len(audiopath_and_texts)))
pass
def dnp(run_name,
noisy_file,
samples_dir,
LR=0.001,
num_iter=5000,
save_every=50):
# initiate model
nlayers = 6
model = Unet(nlayers=nlayers, nefilters=60).cuda()
samples_dir = os.path.join(samples_dir, run_name)
utils.makedirs(samples_dir)
# load data
target, sr = utils.load_wav_to_torch(noisy_file)
target = target[:(len(target) // 2**nlayers) * 2**nlayers]
target = target / utils.MAX_WAV_VALUE
input = torch.rand_like(target)
input = (input - 0.5) * 2
target, input = target.cuda(), input.cuda()
criterion = torch.nn.MSELoss()
optimize(model.parameters(), model, criterion, input, target, samples_dir,
LR, num_iter, sr, save_every)
help='Directory to put Mel-Spectrogram Tensors')
parser.add_argument('-c',
'--config',
type=str,
help='JSON file for configuration')
args = parser.parse_args()
filepaths = utils.files_to_list(args.audio_list)
# Make directory if it doesn't exist
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir)
os.chmod(args.output_dir, 0o775)
# Parse config. Only using data processing
with open(args.config) as f:
data = f.read()
config = json.loads(data)
data_config = config["data_config"]
mel_factory = Mel2SampOnehot(**data_config)
for filepath in filepaths:
audio, sampling_rate = utils.load_wav_to_torch(filepath)
assert (sampling_rate == mel_factory.sampling_rate)
melspectrogram = mel_factory.get_mel(audio)
filename = os.path.basename(filepath)
new_filepath = args.output_dir + '/' + filename + '.pt'
print(new_filepath)
torch.save(melspectrogram, new_filepath)
def get_mel(self, filename):
audio = load_wav_to_torch(filename, self.sampling_rate)
mel = self.mel_transform_fn.transform(audio.view(1, -1))
return mel
hparams = create_hparams()
stft = layers.TacotronSTFT(hparams.filter_length, hparams.hop_length,
hparams.win_length, hparams.n_mel_channels,
hparams.sampling_rate, hparams.mel_fmin,
hparams.mel_fmax)
if gen_mel:
audio_files = sorted(glob('audio/*.wav'))
out_dir = 'mel'
SaveMkdir(out_dir)
for file in tqdm(audio_files):
tqdm.write(file)
file_basename = os.path.basename(file).split('.')[0]
audio_path = os.path.join(hparams.audio_path,
file_basename + '.wav')
audio, sampling_rate = load_wav_to_torch(audio_path,
hparams.sampling_rate)
audio_norm = audio / hparams.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm,
requires_grad=False)
melspec = stft.mel_spectrogram(audio_norm)
#转置存错 即数据行代表帧 列代表特征
melspec = torch.squeeze(melspec, 0).numpy().transpose()
out_file = os.path.join(out_dir, file_basename + '.npy')
np.save(out_file, melspec)
mean_std = cal_MeanStd(out_dir, hparams.n_mel_channels, ref_file=None)
np.save(os.path.join(out_dir, os.pardir, 'MeanStd_Tacotron_mel.npy'),
mean_std)
def __getitem__(self, index):
item = self.audios[index]
return load_wav_to_torch(item)[0]
