Skip to content

Instantly share code, notes, and snippets.

@Sg4Dylan
Last active Jun 18, 2021
Embed
What would you like to do?
「残声」概念验证
#!/usr/bin/env python
#coding:utf-8
# Author: Sg4Dylan --<sg4dylan#gmail.com>
# Created: 06/18/2021
import numpy as np
from scipy import signal
from scipy.io.wavfile import write
import librosa
import resampy
from tqdm import tqdm
# SSB Modulation
def freq_shift_mono(x, f_shift, d_sr):
N_orig = len(x)
N_padded = 2**int(np.ceil(np.log2(np.abs(N_orig))))
S_hilbert = signal.hilbert(np.hstack((x, np.zeros(N_padded-N_orig, x.dtype))))
S_factor = np.exp(2j*np.pi*f_shift*d_sr*np.arange(0, N_padded))
return (S_hilbert*S_factor)[:N_orig].real
def freq_shift_multi(x, f_shift, d_sr):
return np.asarray([freq_shift_mono(x[i], f_shift, d_sr) for i in range(len(x))])
def zansei_impl(x, sr, m=8, decay=1.25):
# Pre-highpass
b,a = signal.butter(11,3000/(sr/2),'high')
d_src = signal.filtfilt(b,a,x)
# !!!
d_sr = 1/sr
f_dn = freq_shift_mono if len(x.shape) == 1 else freq_shift_multi
d_res = np.zeros_like(x)
for i in tqdm(range(m)):
d_res += f_dn(d_src, sr*(i+1)/(m*2), d_sr) * np.exp(-(i+1)*decay)
# Post-highpass
b,a = signal.butter(11,16000/(sr/2),'high')
d_res = signal.filtfilt(b,a,d_res)
# Dynamic-protect
adp_power = np.mean(np.abs(d_res))
src_power = np.mean(np.abs(x))
adj_factor = src_power/(adp_power+src_power)
return (x + d_res)*adj_factor
path_of_audio = 'test.wav'
path_of_result = 'mod.wav'
audio_y, audio_sr = librosa.load(path_of_audio,mono=False,sr=None)
audio_y = resampy.resample(audio_y, audio_sr, audio_sr*2, filter='kaiser_fast')
audio_sr *= 2
shifted = zansei_impl(audio_y, audio_sr)
write(path_of_result, audio_sr, np.asarray(shifted).astype(np.float32).T)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment