mirror of
https://github.com/AIGC-Audio/AudioGPT.git
synced 2025-12-18 20:59:41 +01:00
77 lines
2.1 KiB
Python
77 lines
2.1 KiB
Python
#########
|
|
# world
|
|
##########
|
|
import librosa
|
|
import numpy as np
|
|
import torch
|
|
|
|
gamma = 0
|
|
mcepInput = 3 # 0 for dB, 3 for magnitude
|
|
alpha = 0.45
|
|
en_floor = 10 ** (-80 / 20)
|
|
FFT_SIZE = 2048
|
|
|
|
|
|
f0_bin = 256
|
|
f0_max = 1100.0
|
|
f0_min = 50.0
|
|
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
|
|
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
|
|
|
|
|
|
def f0_to_coarse(f0):
|
|
is_torch = isinstance(f0, torch.Tensor)
|
|
f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700)
|
|
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (f0_mel_max - f0_mel_min) + 1
|
|
|
|
f0_mel[f0_mel <= 1] = 1
|
|
f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1
|
|
f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int)
|
|
assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (f0_coarse.max(), f0_coarse.min())
|
|
return f0_coarse
|
|
|
|
|
|
def norm_f0(f0, uv, hparams):
|
|
is_torch = isinstance(f0, torch.Tensor)
|
|
if hparams['pitch_norm'] == 'standard':
|
|
f0 = (f0 - hparams['f0_mean']) / hparams['f0_std']
|
|
if hparams['pitch_norm'] == 'log':
|
|
f0 = torch.log2(f0) if is_torch else np.log2(f0)
|
|
if uv is not None and hparams['use_uv']:
|
|
f0[uv > 0] = 0
|
|
return f0
|
|
|
|
|
|
def norm_interp_f0(f0, hparams):
|
|
is_torch = isinstance(f0, torch.Tensor)
|
|
if is_torch:
|
|
device = f0.device
|
|
f0 = f0.data.cpu().numpy()
|
|
uv = f0 == 0
|
|
f0 = norm_f0(f0, uv, hparams)
|
|
if sum(uv) == len(f0):
|
|
f0[uv] = 0
|
|
elif sum(uv) > 0:
|
|
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
|
|
uv = torch.FloatTensor(uv)
|
|
f0 = torch.FloatTensor(f0)
|
|
if is_torch:
|
|
f0 = f0.to(device)
|
|
return f0, uv
|
|
|
|
|
|
def denorm_f0(f0, uv, hparams, pitch_padding=None, min=None, max=None):
|
|
if hparams['pitch_norm'] == 'standard':
|
|
f0 = f0 * hparams['f0_std'] + hparams['f0_mean']
|
|
if hparams['pitch_norm'] == 'log':
|
|
f0 = 2 ** f0
|
|
if min is not None:
|
|
f0 = f0.clamp(min=min)
|
|
if max is not None:
|
|
f0 = f0.clamp(max=max)
|
|
if uv is not None and hparams['use_uv']:
|
|
f0[uv > 0] = 0
|
|
if pitch_padding is not None:
|
|
f0[pitch_padding] = 0
|
|
return f0
|