-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathRNNoise_band_energy.py
142 lines (119 loc) · 5.55 KB
/
RNNoise_band_energy.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
132
133
134
135
136
137
138
139
140
141
142
# -*- coding:utf-8 -*-
# Author:凌逆战 | Never.Ling
# Date: 2022/7/15
"""
不是RNNoise原本的刻度，因为RNNoise刻度最高到400，而有481个频点
本文档主要使用mel刻度，但是使用的是RNNoise的打叉法三角滤波器组
RNNoise特征提取和直接和三角滤波器组相乘结果一样
"""
import numpy as np
class MelFilterBank():
def __init__(self, nfilter=32, nfft=512, sr=16000,
lowfreq=0, highfreq=None, transpose=False):
"""
:param nfilter: filterbank中的滤波器数量
:param nfft: FFT size
:param sr: 采样率
:param lowfreq: Mel-filter的最低频带边缘
:param highfreq: Mel-filter的最高频带边缘，默认samplerate/2
"""
self.nfilter = nfilter
self.freq_bins = int(nfft / 2 + 1)
highfreq = highfreq or sr / 2
self.transpose = transpose
# 按梅尔均匀间隔计算 点
lowmel, highmel = self.hz2mel(lowfreq), self.hz2mel(highfreq)
melpoints = np.linspace(lowmel, highmel, nfilter)
hz_points = self.mel2hz(melpoints) # 将mel频率再转到hz频率
# print("hz_points", hz_points)
# bin = (sr/2) / (NFFT+/2) =sample_rate/NFFT # 每个频点的频率数
# bins_list = hz_points/bin=hz_points/(sample_rate/NFFT) # hz_points对应第几个fft频点
self.bins_list = np.floor(hz_points * nfft / sr).astype(np.int32)
print("self.bins_list", len(self.bins_list), self.bins_list) # 32
# [ 0 1 3 5 8 10 13 15 18 22 25 29 33 38 42 48 53 59
# 66 73 80 88 97 107 117 128 140 153 167 182 198 216 235 256]
def hz2mel(self, hz, approach="Oshaghnessy"):
""" Hz to Mels """
return {
"Oshaghnessy": 2595 * np.log10(1 + hz / 700.0),
"Lindsay": 2410 * np.log10(1 + hz / 625),
}[approach]
def mel2hz(self, mel, approach="Oshaghnessy"):
""" Mels to HZ """
return {
"Oshaghnessy": 700 * (10 ** (mel / 2595.0) - 1),
"Lindsay": 625 * (10 ** (mel / 2410) - 1),
}[approach]
def get_filter_bank(self):
# RNNoise的三角滤波器，打叉画法
fbank = np.zeros((self.nfilter, self.freq_bins)) # (M,F)
for i in range(self.nfilter - 1):
band_size = (self.bins_list[i + 1] - self.bins_list[i])
for j in range(band_size):
frac = j / band_size
fbank[i, self.bins_list[i] + j] = 1 - frac # 降
fbank[i + 1, self.bins_list[i] + j] = frac # 升
# 第一个band和最后一个band的窗只有一半因而能量乘以2
fbank[0] *= 2
fbank[-1] *= 2
return fbank
def interp_band_gain(self, gain):
# gain (M,T)
gain_interp = np.zeros((self.freq_bins, gain.shape[-1])) # (F,T)
for i in range(self.nfilter - 1):
band_size = (self.bins_list[i + 1] - self.bins_list[i])
for j in range(band_size):
frac = j / band_size
# gain_interp[eband5ms[i] * 4 + j] = frac * gain[i] + \
# (1 - frac) * gain[i + 1]
gain_interp[self.bins_list[i] + j] = (1 - frac) * gain[i] + \
frac * gain[i + 1]
return gain_interp
if __name__ == '__main__':
import matplotlib.pyplot as plt
import librosa.display
from matplotlib.ticker import FuncFormatter
def formatnum(x, pos):
return '$%d$' % (x / 1000)
formatter = FuncFormatter(formatnum)
sr = 16000
band_nums = 32
NFFT = win_length = 512
freq_bins = NFFT // 2 + 1
eband5ms = (0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 20, 24, 28, 34, 40, 48, 60, 78, 100)
x = librosa.load("../wav_data/clean/TIMIT/TRAIN_DR1_FCJF0_SA1.wav", sr=sr)[0]
X = librosa.stft(x, n_fft=NFFT, hop_length=win_length // 2,
win_length=win_length,
window="hann", center=False) # (F,T)
mag = np.abs(X) # (F,T)
FBank_class = MelFilterBank(nfilter=band_nums, nfft=NFFT,
sr=sr, lowfreq=0, highfreq=None, transpose=False)
Fbank = FBank_class.get_filter_bank() # (M,F)
filter_banks = np.dot(Fbank, mag) # (M,F) * (F,T) = (M,T)
# ================ 绘制三角滤波器组 ==========================
fig = plt.figure(figsize=(8, 12))
fig.suptitle('RNNoise_fbank_nfft={}_filter_nums={}'.format(NFFT, band_nums))
plt.subplot(3, 1, 1)
fs_bin = sr // 2 / (NFFT // 2) # 一个频点多少Hz
x = np.linspace(0, freq_bins, freq_bins)
plt.plot(x, Fbank.T)
plt.subplot(3, 1, 2)
librosa.display.specshow(20 * np.log10(filter_banks), sr=sr, x_axis='time', y_axis='linear', cmap="jet")
plt.xlabel('time/s', fontsize=14)
plt.ylabel('freqs/kHz', fontsize=14)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.gca().yaxis.set_major_formatter(formatter)
plt.subplot(3, 1, 3)
interp_band_gain = FBank_class.interp_band_gain(filter_banks) # (F,T)
print("插值结果：", interp_band_gain[:10, 1])
transpose_band_gain = np.dot(Fbank.T, filter_banks) # (F,M)*(M*T)=(F,T)
print("转置相乘：", transpose_band_gain[:10, 1])
librosa.display.specshow(20 * np.log10(interp_band_gain), sr=sr, x_axis='time', y_axis='linear', cmap="jet")
plt.xlabel('Time/s', fontsize=14)
plt.ylabel('Freqs/kHz', fontsize=14)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.gca().yaxis.set_major_formatter(formatter)
plt.tight_layout()
plt.show()