forked from NVIDIA/mellotron
-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_utils.py
executable file
·160 lines (133 loc) · 6.1 KB
/
data_utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
import random
import os
import re
import numpy as np
import torch
import torch.utils.data
import librosa
import layers
from utils import load_wav_to_torch, load_filepaths_and_text
from text import text_to_sequence, cmudict
from yin import compute_yin
class TextMelLoader(torch.utils.data.Dataset):
"""
1) loads audio, text and speaker ids
2) normalizes text and converts them to sequences of one-hot vectors
3) computes mel-spectrograms and f0s from audio files.
"""
def __init__(self, audiopaths_and_text, hparams, speaker_ids=None):
self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
self.text_cleaners = hparams.text_cleaners
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.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)
self.sampling_rate = hparams.sampling_rate
self.filter_length = hparams.filter_length
self.hop_length = hparams.hop_length
self.f0_min = hparams.f0_min
self.f0_max = hparams.f0_max
self.harm_thresh = hparams.harm_thresh
self.p_arpabet = hparams.p_arpabet
self.cmudict = None
if hparams.cmudict_path is not None:
self.cmudict = cmudict.CMUDict(hparams.cmudict_path)
self.speaker_ids = speaker_ids
if speaker_ids is None:
self.speaker_ids = self.create_speaker_lookup_table(
self.audiopaths_and_text)
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def create_speaker_lookup_table(self, audiopaths_and_text):
speaker_ids = np.sort(np.unique([x[2] for x in audiopaths_and_text]))
d = {int(speaker_ids[i]): i for i in range(len(speaker_ids))}
return d
def get_f0(self, audio, sampling_rate=22050, frame_length=1024,
hop_length=256, f0_min=100, f0_max=300, harm_thresh=0.1):
f0, harmonic_rates, argmins, times = compute_yin(
audio, sampling_rate, frame_length, hop_length, f0_min, f0_max,
harm_thresh)
pad = int((frame_length / hop_length) / 2)
f0 = [0.0] * pad + f0 + [0.0] * pad
f0 = np.array(f0, dtype=np.float32)
return f0
def get_data(self, audiopath_and_text):
audiopath, text, speaker = audiopath_and_text
text = self.get_text(text)
mel, f0 = self.get_mel_and_f0(audiopath)
speaker_id = self.get_speaker_id(speaker)
return (text, mel, speaker_id, f0)
def get_speaker_id(self, speaker_id):
return torch.IntTensor([self.speaker_ids[int(speaker_id)]])
def get_mel_and_f0(self, filepath):
audio, sampling_rate = load_wav_to_torch(filepath)
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)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
f0 = self.get_f0(audio.cpu().numpy(), self.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)]
return melspec, f0
def get_text(self, text):
text_norm = torch.IntTensor(
text_to_sequence(text, self.text_cleaners, self.cmudict, self.p_arpabet))
return text_norm
def __getitem__(self, index):
return self.get_data(self.audiopaths_and_text[index])
def __len__(self):
return len(self.audiopaths_and_text)
class TextMelCollate():
""" Zero-pads model inputs and targets based on number of frames per setep
"""
def __init__(self, n_frames_per_step):
self.n_frames_per_step = n_frames_per_step
def __call__(self, batch):
"""Collate's training batch from normalized text and mel-spectrogram
PARAMS
------
batch: [text_normalized, mel_normalized]
"""
# Right zero-pad all one-hot text sequences to max input length
input_lengths, ids_sorted_decreasing = torch.sort(
torch.LongTensor([len(x[0]) for x in batch]),
dim=0, descending=True)
max_input_len = input_lengths[0]
text_padded = torch.LongTensor(len(batch), max_input_len)
text_padded.zero_()
for i in range(len(ids_sorted_decreasing)):
text = batch[ids_sorted_decreasing[i]][0]
text_padded[i, :text.size(0)] = text
# Right zero-pad mel-spec
num_mels = batch[0][1].size(0)
max_target_len = max([x[1].size(1) for x in batch])
if max_target_len % self.n_frames_per_step != 0:
max_target_len += self.n_frames_per_step - max_target_len % self.n_frames_per_step
assert max_target_len % self.n_frames_per_step == 0
# include mel padded, gate padded and speaker ids
mel_padded = torch.FloatTensor(len(batch), num_mels, max_target_len)
mel_padded.zero_()
gate_padded = torch.FloatTensor(len(batch), max_target_len)
gate_padded.zero_()
output_lengths = torch.LongTensor(len(batch))
speaker_ids = torch.LongTensor(len(batch))
f0_padded = torch.FloatTensor(len(batch), 1, max_target_len)
f0_padded.zero_()
for i in range(len(ids_sorted_decreasing)):
mel = batch[ids_sorted_decreasing[i]][1]
mel_padded[i, :, :mel.size(1)] = mel
gate_padded[i, mel.size(1)-1:] = 1
output_lengths[i] = mel.size(1)
speaker_ids[i] = batch[ids_sorted_decreasing[i]][2]
f0 = batch[ids_sorted_decreasing[i]][3]
f0_padded[i, :, :f0.size(1)] = f0
model_inputs = (text_padded, input_lengths, mel_padded, gate_padded,
output_lengths, speaker_ids, f0_padded)
return model_inputs