forked from jabeim/hackathon_vocoder
-
Notifications
You must be signed in to change notification settings - Fork 0
/
vocoder_tools.py
131 lines (111 loc) · 4.72 KB
/
vocoder_tools.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
# -*- coding: utf-8 -*-
"""
Created on Thu May 9 14:32:18 2019
@author: Jbeim
"""
import numpy as np
import scipy as sp
from numba import jit
#@jit('float64[:,:](float64[:,:],float64[:,:],float64[:,:],float64[:,:],float64[:,:],float64,float64,int32,int32,int32,int32)',nopython=True)
#def NeurToAudio(elData,normRamp,normOffset,mNeurToBin,audioPwr,alpha,mAvg,nAvg,playOverAvgRatio,blkSize,nFFT):
## audioPwr = np.zeros((nNeuralLocs,blkSize+1))
#
# shLen = nFFT+(playOverAvgRatio-1)*nAvg
# stateHolder = np.zeros(shLen)
# shli = np.int(0)
# sH2 = np.zeros(shLen)
# nl = 5
# phs = 2*np.pi*np.random.rand(np.floor(nFFT/2).astype(int32))
# dphi = 2*np.pi*np.arange(1,np.floor(nFFT/2)+1)*nAvg/nFFT
# audioOut = np.array([])
#
# for blkNumber in range(1,np.floor(elData.shape[1]/blkSize.astype(int))+1):
# timeIdx = np.arange((blkNumber-1)*blkSize+1,blkNumber*blkSize+1,dtype=int)-1
# efData = np.dot(normRamp,elData[:,timeIdx])
# efData = (efData.T-normOffset).T
# electricField = np.maximum(0,efData)
# activity = np.maximum(0,np.minimum(np.exp(-nl+nl*electricField),1)-np.exp(-nl))/(1-np.exp(-nl))
# audioPwr = ActivityToPower(alpha,activity,audioPwr,blkSize)
# energy = np.sum(audioPwr,axis = 1)/mAvg
# spect = np.multiply(np.dot(mNeurToBin,energy),np.exp(1j*phs))
#
# scl = 1
# if np.mod(nFFT,2) == 1:
# sgn = np.multiply(scl*(nFFT/2)*win,np.real(np.fft.ifft(np.concatenate((np.array([0]),spect,np.conj(spect[::-1]))))))
# else:
# sgn = np.multiply(scl*(nFFT/2)*win,np.real(np.fft.ifft(np.concatenate((np.array([0]),spect,np.conj(spect[spect.size-2::-1]))))))
#
# shWin = np.arange(shli,shli+nFFT)
# stateHolder[shWin] = stateHolder[shWin]+sgn
# shli = shli+nAvg
# phs = np.mod(phs+dphi,2*np.pi)
# shWin2 = np.arange(0,playOverAvgRatio*nAvg)
# if not(np.mod(blkNumber,playOverAvgRatio)):
# audioOut = np.append(audioOut,1*stateHolder[shWin2])
# stateHolder = np.concatenate((stateHolder[playOverAvgRatio*nAvg:None],np.zeros(playOverAvgRatio*nAvg)))
# sH2 = np.concatenate((sH2[playOverAvgRatio*nAvg:None],np.zeros(playOverAvgRatio*nAvg)))
# shli = np.int(0)
#
#
#
#
# return audioOut
@jit('float64[:,:](float64,float64[:,:],float64[:,:],int32)',nopython = True)
def ActivityToPower(alpha,activity,audioPwr,blkSize):
for k in range(blkSize):
audioPwr[:,k+1] = np.maximum(audioPwr[:,k]*alpha+activity[:,k]*(1-alpha),activity[:,k])
audioPwr[:,0] = audioPwr[:,blkSize]
return audioPwr
def NeurToBinMatrix(neuralLocsOct,nFFT,Fs):
fGrid = np.arange(0,np.floor(nFFT/2)+1)*(Fs/nFFT)
fBinsOct = np.log2(fGrid[1:])
binCountPerOct = np.divide(1,np.diff(fBinsOct))
x = np.ones((2,np.floor(nFFT/2).astype(int)-1))
x[1,:] = np.arange(1,np.floor(nFFT/2),1)
## x = np.append(a1,a2,axis=1)
# coef = np.linalg.solve(x,binCountPerOct)
# scl = coef[0]+coef[1]*10**(neuralLocsOct/20)
nNeuralLocs = len(neuralLocsOct)
mNeurToBin = np.zeros((np.floor(nFFT/2).astype(int),nNeuralLocs))
I = np.zeros(nNeuralLocs)
for k in range(len(neuralLocsOct)):
tmp = np.abs(fBinsOct-neuralLocsOct[k])
I[k] = np.argmin(tmp)
mNeurToBin[I[k].astype(int),k] = 1
pFN = 'preemph.mat'
emph = sp.io.loadmat(pFN)
I = np.argmax(emph['emphDb'])
emph['emphDb'][I+1:] = emph['emphDb'][I]
emphDb = -emph['emphDb']
emphDb= emphDb-emphDb[0]
scl = np.interp(
np.arange(1,np.floor(nFFT/2)+1)*Fs/nFFT,
np.append(0,emph['emphF']),
np.append(0,emphDb)
)
mNeurToBin = np.multiply(mNeurToBin.T,10**(scl/20))
mNeurToBin = np.nan_to_num(mNeurToBin).T
return mNeurToBin
def generate_cfs(lo, hi, n_bands):
"""
Generates a series of 'bands' frequencies in Hz, linearely distributed
on an ERB scale between the frequencies 'lo' and 'hi' (in Hz).
These would are the centre frequencies (on an ERB scale) of the bands
specifications made by 'generate_bands' with the same arguments
"""
density = n_bands / (hz2erb(hi) - hz2erb(lo))
bands = []
for i in np.arange(1, n_bands + 1):
bands.append(erb2hz(hz2erb(lo) + (i - 0.5) / density))
return bands
def erb2hz(erb):
"""
Convert equivalent rectangular bandwidth (ERB) to Hertz.
"""
tmp = np.exp((erb - 43.) / 11.17)
return (0.312 - 14.675 * tmp) / (tmp - 1.0) * 1000.
def hz2erb(hz):
"""
Convert Hertz to equivalent rectangular bandwidth (ERB).
"""
return 11.17 * np.log((hz + 312.) / (hz + 14675.)) + 43.