-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdata.py
142 lines (124 loc) · 4.31 KB
/
data.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
"""
Create tensorflow dataset
"""
import os
import random
from argparse import ArgumentParser
from pathlib import Path
import jax
import jax.numpy as jnp
import librosa
import numpy as np
import tensorflow as tf
from tqdm.cli import tqdm
from dsp import MelFilter
from utils import get_wav_files, load_config
config = load_config()
SAMPLE_RATE = config["SAMPLE_RATE"]
MEL_DIM = config["MEL_DIM"]
MEL_MIN = config["MEL_MIN"]
PAD_TOKEN = config["PAD_TOKEN"]
PAD = config["PAD"]
END_CHARACTER = config["END_CHARACTER"]
TF_DATA_DIR = Path(config["TF_DATA_DIR"])
SAMPLE_RATE = config["SAMPLE_RATE"]
WINDOW_LENGTH = SAMPLE_RATE * 500 // 10_000 # 50.0 ms
HOP_LENGTH = SAMPLE_RATE * 125 // 10_000 # 12.5 ms
# Extract log-melspectrogram features from a wav file
# - window size: 50 ms
# - hop length : 12.5 ms
mel_filter = MelFilter(
sample_rate=SAMPLE_RATE,
n_fft=2048,
window_length=WINDOW_LENGTH,
hop_length=HOP_LENGTH,
n_mels=MEL_DIM,
fmin=0,
fmax=SAMPLE_RATE // 2,
mel_min=MEL_MIN,
)
mel_filter = jax.jit(mel_filter)
def get_transcripts_w_end_character(wav_files):
"""
Read all *.txt files that corresponding to *.wav files.
Add END_CHARACTER to the end of the transcripts.
"""
texts = []
for file_path in wav_files:
txt_path = file_path.with_suffix(".txt")
with open(txt_path, "r", encoding="utf-8") as f:
text = f.read().strip()
assert PAD not in text
assert END_CHARACTER not in text
text = text + END_CHARACTER
texts.append(text)
return texts
def get_alphabet(wav_files):
"""
Return a list of all characters in the transcripts.
The [pad] character and [end] character are also included.
"""
texts = get_transcripts_w_end_character(wav_files)
alphabet = sorted(set("".join(texts)) - {END_CHARACTER})
return [PAD, END_CHARACTER] + alphabet
def create_tf_data(data_dir: Path, output_dir: Path):
"""
Create a tensorflow dataset
"""
wav_files = get_wav_files(data_dir)
texts = get_transcripts_w_end_character(wav_files)
maxlen = max(map(len, texts))
padded_texts = [l + PAD * (maxlen - len(l)) for l in texts]
alphabet = get_alphabet(wav_files)
assert PAD_TOKEN == alphabet.index(PAD)
text_tokens = []
for text in tqdm(padded_texts):
o = []
for c in text:
o.append(alphabet.index(c))
text_tokens.append(o)
sorted_files = sorted(wav_files, key=os.path.getsize)
y, _ = librosa.load(sorted_files[-1], sr=SAMPLE_RATE, res_type="soxr_hq")
mel = mel_filter(y[None])[0].astype(jnp.float16)
mel_shape = mel.shape
max_wav_len = len(y)
def data_generator():
data = list(zip(text_tokens, wav_files))
random.Random(42).shuffle(data)
data = tqdm(data)
for text, wav_file in data:
wav, rate = librosa.load(wav_file, sr=SAMPLE_RATE, res_type="soxr_hq")
assert rate == SAMPLE_RATE
assert max_wav_len >= len(wav)
wav = wav / max(1.0, np.max(np.abs(wav))) # rescale to avoid overflow
pads = [(0, max_wav_len - wav.shape[0])]
mel_len = len(wav) // HOP_LENGTH + 1
wav = np.pad(wav, pads, constant_values=0)
mel = mel_filter(wav[None])[0].astype(jnp.float16)
mel = mel.at[mel_len:].set(0)
mel = jax.device_get(mel)
text = np.array(text, dtype=np.int32)
yield wav_file.stem, text, mel
output_signature = (
tf.TensorSpec(shape=(), dtype=tf.string),
tf.TensorSpec(shape=[len(text_tokens[0])], dtype=tf.int32),
tf.TensorSpec(shape=mel_shape, dtype=tf.float16),
)
dataset = tf.data.Dataset.from_generator(
data_generator, output_signature=output_signature
)
tf.data.experimental.save(dataset, str(output_dir))
# save the alphabet for inference mode
with open(output_dir / "alphabet.txt", "w", encoding="utf-8") as file:
for ch in alphabet:
file.write(ch + "\n")
parser = ArgumentParser()
parser.add_argument("wav_dir", type=Path)
wav_dir = parser.parse_args().wav_dir
# prepare tensorflow dataset
print(f"Loading data from directory '{wav_dir}'")
create_tf_data(wav_dir, TF_DATA_DIR)
print(
f"Created a tensorflow dataset at '{TF_DATA_DIR}'.\n\n"
f"Run 'python train.py' to train your model."
)