File size: 3,765 Bytes
5e09ea1 |
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 |
import os
import argparse
import torch
import json
from glob import glob
from pyworld import pyworld
from tqdm import tqdm
from scipy.io import wavfile
import utils
from mel_processing import mel_spectrogram_torch
#import h5py
import logging
logging.getLogger('numba').setLevel(logging.WARNING)
import parselmouth
import librosa
import numpy as np
def stft(y):
return librosa.stft(
y=y,
n_fft=1280,
hop_length=320,
win_length=1280,
)
def energy(y):
# Extract energy
S = librosa.magphase(stft(y))[0]
e = np.sqrt(np.sum(S ** 2, axis=0)) # np.linalg.norm(S, axis=0)
return e.squeeze() # (Number of frames) => (654,)
def get_energy(path, p_len=None):
wav, sr = librosa.load(path, 48000)
e = energy(wav)
if p_len is None:
p_len = wav.shape[0] // 320
assert e.shape[0] -p_len <2 ,(e.shape[0] ,p_len)
e = e[: p_len]
return e
def get_f0(path,p_len=None, f0_up_key=0):
x, _ = librosa.load(path, 48000)
if p_len is None:
p_len = x.shape[0]//320
else:
assert abs(p_len-x.shape[0]//320) < 3, (path, p_len, x.shape)
time_step = 320 / 48000 * 1000
f0_min = 50
f0_max = 1100
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
f0 = parselmouth.Sound(x, 48000).to_pitch_ac(
time_step=time_step / 1000, voicing_threshold=0.6,
pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
pad_size=(p_len - len(f0) + 1) // 2
if(pad_size>0 or p_len - len(f0) - pad_size>0):
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
f0bak = f0.copy()
f0 *= pow(2, f0_up_key / 12)
f0_mel = 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > 255] = 255
f0_coarse = np.rint(f0_mel).astype(np.int)
return f0_coarse, f0bak
def resize2d(x, target_len):
source = np.array(x)
source[source<0.001] = np.nan
target = np.interp(np.arange(0, len(source)*target_len, len(source))/ target_len, np.arange(0, len(source)), source)
res = np.nan_to_num(target)
return res
def compute_f0(path, c_len):
x, sr = librosa.load(path, sr=48000)
f0, t = pyworld.dio(
x.astype(np.double),
fs=sr,
f0_ceil=800,
frame_period=1000 * 320 / sr,
)
f0 = pyworld.stonemask(x.astype(np.double), f0, t, 48000)
for index, pitch in enumerate(f0):
f0[index] = round(pitch, 1)
assert abs(c_len - x.shape[0]//320) < 3, (c_len, f0.shape)
return None, resize2d(f0, c_len)
def process(filename):
print(filename)
save_name = filename+".soft.pt"
if not os.path.exists(save_name):
devive = torch.device("cuda" if torch.cuda.is_available() else "cpu")
wav, _ = librosa.load(filename, sr=16000)
wav = torch.from_numpy(wav).unsqueeze(0).to(devive)
c = utils.get_hubert_content(hmodel, wav)
torch.save(c.cpu(), save_name)
else:
c = torch.load(save_name)
f0path = filename+".f0.npy"
if not os.path.exists(f0path):
cf0, f0 = compute_f0(filename, c.shape[-1] * 3)
np.save(f0path, f0)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--in_dir", type=str, default="dataset/48k", help="path to input dir")
args = parser.parse_args()
print("Loading hubert for content...")
hmodel = utils.get_hubert_model(0 if torch.cuda.is_available() else None)
print("Loaded hubert.")
filenames = glob(f'{args.in_dir}/*/*.wav', recursive=True)#[:10]
for filename in tqdm(filenames):
process(filename)
|