Spaces:
Sleeping
Sleeping
File size: 2,804 Bytes
c42fe7e |
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 |
import os
import sys
import librosa
import numpy as np
import resampy
import torch
import torchcrepe
import tqdm
from utils.binarizer_utils import get_pitch_parselmouth, get_mel_torch
from modules.vocoders.nsf_hifigan import NsfHifiGAN
from utils.infer_utils import save_wav
from utils.hparams import set_hparams, hparams
sys.argv = [
'inference/svs/ds_acoustic.py',
'--config',
'configs/acoustic.yaml',
]
def get_pitch(wav_data, mel, hparams, threshold=0.3):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# crepe只支持16khz采样率，需要重采样
wav16k = resampy.resample(wav_data, hparams['audio_sample_rate'], 16000)
wav16k_torch = torch.FloatTensor(wav16k).unsqueeze(0).to(device)
# 频率范围
f0_min = 40
f0_max = 800
# 重采样后按照hopsize=80,也就是5ms一帧分析f0
f0, pd = torchcrepe.predict(wav16k_torch, 16000, 80, f0_min, f0_max, pad=True, model='full', batch_size=1024,
device=device, return_periodicity=True)
# 滤波，去掉静音，设置uv阈值，参考原仓库readme
pd = torchcrepe.filter.median(pd, 3)
pd = torchcrepe.threshold.Silence(-60.)(pd, wav16k_torch, 16000, 80)
f0 = torchcrepe.threshold.At(threshold)(f0, pd)
f0 = torchcrepe.filter.mean(f0, 3)
# 将nan频率（uv部分）转换为0频率
f0 = torch.where(torch.isnan(f0), torch.full_like(f0, 0), f0)
# 去掉0频率，并线性插值
nzindex = torch.nonzero(f0[0]).squeeze()
f0 = torch.index_select(f0[0], dim=0, index=nzindex).cpu().numpy()
time_org = 0.005 * nzindex.cpu().numpy()
time_frame = np.arange(len(mel)) * hparams['hop_size'] / hparams['audio_sample_rate']
f0 = np.interp(time_frame, time_org, f0, left=f0[0], right=f0[-1])
return f0
set_hparams()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
vocoder = NsfHifiGAN()
in_path = 'path/to/input/wavs'
out_path = 'path/to/output/wavs'
os.makedirs(out_path, exist_ok=True)
for filename in tqdm.tqdm(os.listdir(in_path)):
if not filename.endswith('.wav'):
continue
wav, _ = librosa.load(os.path.join(in_path, filename), sr=hparams['audio_sample_rate'], mono=True)
mel = get_mel_torch(
wav, hparams['audio_sample_rate'], num_mel_bins=hparams['audio_num_mel_bins'],
hop_size=hparams['hop_size'], win_size=hparams['win_size'], fft_size=hparams['fft_size'],
fmin=hparams['fmin'], fmax=hparams['fmax'],
device=device
)
f0, _ = get_pitch_parselmouth(
wav, samplerate=hparams['audio_sample_rate'], length=len(mel),
hop_size=hparams['hop_size']
)
wav_out = vocoder.spec2wav(mel, f0=f0)
save_wav(wav_out, os.path.join(out_path, filename), hparams['audio_sample_rate'])
|