def __init__(self, training_files, segment_length, filter_length,
hop_length, win_length, sampling_rate, mel_fmin, mel_fmax):
if os.path.isfile(str(training_files)):
self.audio_files = files_to_list(training_files)
else:
self.audio_files = []
random.seed(1234)
random.shuffle(self.audio_files)
self.stft = TacotronSTFT(filter_length=filter_length,
hop_length=hop_length,
win_length=win_length,
sampling_rate=sampling_rate,
mel_fmin=mel_fmin, mel_fmax=mel_fmax)
self.segment_length = segment_length
self.sampling_rate = sampling_rate
class Mel2Samp(torch.utils.data.Dataset):
"""
This is the main class that calculates the spectrogram and returns the
spectrogram, audio pair.
"""
def __init__(self, training_files, segment_length, filter_length,
hop_length, win_length, sampling_rate, mel_fmin, mel_fmax):
if os.path.isfile(str(training_files)):
self.audio_files = files_to_list(training_files)
else:
self.audio_files = []
random.seed(1234)
random.shuffle(self.audio_files)
self.stft = TacotronSTFT(filter_length=filter_length,
hop_length=hop_length,
win_length=win_length,
sampling_rate=sampling_rate,
mel_fmin=mel_fmin, mel_fmax=mel_fmax)
self.segment_length = segment_length
self.sampling_rate = sampling_rate
def get_mel(self, audio):
audio_norm = audio # audio / 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)
return melspec
def __getitem__(self, index):
# Read audio
filename = self.audio_files[index]
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))
# 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 = audio / MAX_WAV_VALUE
return (mel, audio)
def __len__(self):
return len(self.audio_files)
import unidecode
import yaml
import librosa
from waveglow import inference as waveglow
from melgan import inference as melgan
from mellotron import inference as mellotron
from utils.argutils import locals2dict
from mellotron.layers import TacotronSTFT
from mellotron.hparams import create_hparams
# 用griffinlim声码器
_hparams = create_hparams()
_stft = TacotronSTFT(_hparams.filter_length, _hparams.hop_length,
_hparams.win_length, _hparams.n_mel_channels,
_hparams.sampling_rate, _hparams.mel_fmin,
_hparams.mel_fmax)
_use_waveglow = 0
_device = 'cuda' if torch.cuda.is_available() else 'cpu'
filename_formatter_re = re.compile(r'[\s\\/:*?"<>|\']+')
def plot_mel_alignment_gate_audio(mel,
alignment,
gate,
audio,
figsize=(16, 16)):
fig, axes = plt.subplots(4, 1, figsize=figsize)
axes = axes.flatten()
import torch
import aukit
import tqdm
import requests
from waveglow import inference as waveglow
from mellotron import inference as mellotron
from mellotron.layers import TacotronSTFT
from mellotron.hparams import create_hparams
_home_dir = os.path.dirname(os.path.abspath(__file__))
# 用griffinlim声码器
_hparams = create_hparams()
_stft = TacotronSTFT(_hparams.filter_length, _hparams.hop_length,
_hparams.win_length, _hparams.n_mel_channels,
_hparams.sampling_rate, _hparams.mel_fmin,
_hparams.mel_fmax)
_use_waveglow = 0
_device = 'cuda' if torch.cuda.is_available() else 'cpu'
_mellotron_path = os.path.join(_home_dir, 'resource', 'model',
'mellotron.kuangdd-rtvc.pt')
_waveglow_path = os.path.join(_home_dir, 'resource', 'model',
'waveglow.kuangdd.pt')
_ge2e_path = os.path.join(_home_dir, 'resource', 'model', 'ge2e.kuangdd.pt')
_mellotron_hparams_path = os.path.join(_home_dir, 'resource', 'model',
'mellotron_hparams.json')
_reference_audio_tar_path = os.path.join(_home_dir, 'resource',
'reference_audio.tar')
_audio_tar_path = os.path.join(_home_dir, 'resource', 'audio.tar')