tokenpocket官网苹果下载|gfcc

首页
gfcc

tokenpocket官网苹果下载|gfcc

作者： tokenpocket官网苹果下载

2024-03-17 09:48:18

Global Federation of Competitiveness Councils

top of pagePressNewsletterThe GFCCHomeAbout the GFCCWhat the GFCC doesAnnual ReportsGFCC MembersGFCC FellowsThought LeadershipBlogGlobal Competitiveness PrinciplesCall to ActionUniversity and InnovationAdditional ReportsInitiativesDriving Innovation in Times of CrisisUniversity & Research Leadearship ForumGlobal Competitiveness AwardCompetitiveness DecoderFrame the Future SeriesEventsNews and AnnouncementsGFCC CommunityExclusive CommunityGFCC StoriesMoreUse tab to navigate through the menu items.PressNewsletterThe GFCCHomeAbout the GFCCWhat the GFCC doesAnnual ReportsGFCC MembersGFCC FellowsThought LeadershipBlogGlobal Competitiveness PrinciplesCall to ActionUniversity and InnovationAdditional ReportsInitiativesDriving Innovation in Times of CrisisUniversity & Research Leadearship ForumGlobal Competitiveness AwardCompetitiveness DecoderFrame the Future SeriesEventsNews and AnnouncementsGFCC CommunityExclusive CommunityGFCC StoriesMoreUse tab to navigate through the menu items.Welcome to the

Global Federation of Competitiveness CouncilsA global

multi-stakeholder membership organization founded in 2010The GFCC is committed to disseminating best practices to accelerate productivity, growth, and prosperity for countries, regions, and cities.

The GFCC does that through high-level networking and events, in-depth conversations, advice, and education.

Learn more about GFCC's purpose and offerThe GFCC provides exposure and access to a unique network of global leaders, innovation agencies, corporations, and research institutes through its networking activities and knowledge-sharing platform.Read more about the GFCC's DNA

InitiativesDriving Innovation in Times of CrisisDriving Innovation in Times of Crisis deepens the understanding of crises to unlock new opportunities for innovation, enhance competitiveness, and promote resilience.

Read MoreUniversity and Research Leadership ForumThe University and Research Leadership Forum is a living think-tank to examine trends and generates ideas to optimize education and research enterprises.

Read MoreGlobal Competitiveness AwardThe Global Competitiveness Award honors leaders who have advanced the competitiveness agenda.

Read MoreInnovation Learning LabsInnovation Learning Labs are workshops designed to spark progressive dialogues and enhance knowledge of innovation policies and initiatives.

Read MoreThought LeadershipUniversity and InnovationThese reports showcase the pioneering work of the GFCC and its community at the intersection of innovation and higher education.

Read all GFCC reports hereCall to ActionThese reports give recommendations to stakeholders interested in promoting innovation and competitiveness worldwide.

Read all GFCC Call to Actions hereGFCC BlogCompetitive Edge serves as a platform for our community to share knowledge, expose innovative ideas, and learn about the newest innovation trends.

Access the GFCC blog hereUpcoming EventsGlobal Innovation Summit 2024

The GFCC 2024 Annual Meeting and Global Innovation Summit will take place in Belfast from November 11th to 15th, 2024, and will be hosted by Queen's University of Belfast and the Centre for Competitiveness.

See more.In Person and Virtual EventsThe GFCC curates an exclusive collection of both online and in-person events.

Often implemented in partnership with member organizations, they gather members, fellows, and partners from around the globe.

See all GFCC eventsNews & AnnouncementsUniversities in Action: Building Cross-sector Alliances and Making Impact on SocietyRead MoreGet in touch with the GFCC

THE GFCC

Homepage

About the GFCC

What the GFCC does

GFCC Members

GFCC Fellows

THOUGHT LEADERSHIP

Blog

Annual Reports

Annual Principles

Thought Pieces

University & Innovation

INITIATIVES

Driving Innovation in Time of Crises

University & Research Leadearship Forum

Global Competitiveness Award

Frame the Future Series

Competitiveness Decoder

EVENTS

All Events

Past In Person Events

Past Online Events

NEWS & ANNOUNCEMENTS

GFCC COMMUNITY

Global Federation of Competitiveness

Councils

900 17th Street NW, Suite 700

Washington, D.C. 20006

+1 (202) 969 3382

info@thegfcc.orgbottom of page

声学感知刻度(mel scale、Bark scale、ERB)与声学特征提取(MFCC、BFCC、GFCC)_hz2erb-CSDN博客

声学感知刻度(mel scale、Bark scale、ERB)与声学特征提取(MFCC、BFCC、GFCC)

凌逆战

已于 2022-06-16 21:52:28 修改

阅读量5.1k

点赞数

文章标签：

python

开发语言

于 2022-05-28 19:42:00 首次发布

CSDN的所有文章均转载自我博客园的文章，由于存在转载丢失，想了解细节，可访问我的博客园。 https://www.cnblogs.com/LXP-Never/

本文链接：https://blog.csdn.net/qq_34218078/article/details/125145458

版权

本文地址：声学感知刻度(mel scale、Bark scale、ERB)与声学特征提取(MFCC、BFCC、GFCC) - 凌逆战 - 博客园 (引用请注明出处)

本文代码：GitHub - LXP-Never/perception_scale: Human ear perception scales and feature（mel、bark、ERB、gammatone）

作者：凌逆战 | Never.Ling

梅尔刻度

　　梅尔刻度(Mel scale)是一种由听众判断不同频率音高(pitch)彼此相等的感知刻度，表示人耳对等距音高(pitch)变化的感知。mel 刻度和正常频率(Hz)之间的参考点是将1 kHz，且高于人耳听阈值40分贝以上的基音，定为1000 mel。在大约500 Hz以上，听者判断越来越大的音程(interval)产生相等的pitch增量，人耳每感觉到等量的音高变化，所需要的频率变化随频率增加而愈来愈大。

　　将频率$f$ (Hz)转换为梅尔$m$的公式是：

$$m=2595\log_{10}(1+\frac{f}{700})$$

def hz2mel(hz):

""" Hz to Mels """

return 2595 * np.log10(1 + hz / 700.0)

mel与f(Hz)的对应关系

import numpy as np

import matplotlib.pyplot as plt

from matplotlib.ticker import FuncFormatter

def hz2mel(hz):

""" Hz to Mels """

return 2595 * np.log10(1 + hz / 700.0)

if __name__ == "__main__":

fs = 16000

hz = np.linspace(0, 8000, 8000)

mel = hz2mel(hz)

fig = plt.figure()

ax = plt.plot(hz, mel, color="r")

plt.xlabel("Hertz scale (Hz)", fontsize=12) # x轴的名字

plt.ylabel("mel scale", fontsize=12)

plt.xticks(fontsize=10) # x轴的刻度

plt.yticks(fontsize=10)

plt.xlim(0, 8000) # 坐标轴的范围

plt.ylim(0)

def formatnum(x, pos):

return '$%.1f$' % (x / 1000)

formatter = FuncFormatter(formatnum)

# plt.gca().xaxis.set_major_formatter(formatter)

# plt.gca().yaxis.set_major_formatter(formatter)

plt.grid(linestyle='--')

plt.tight_layout()

plt.show()

画图代码

将梅尔$m$转换为频率$f$ (Hz)的公式是：

$$f=700e^{\frac{m}{2595}-1}$$

def mel2hz(mel):

""" Mels to HZ """

return 700 * (10 ** (mel / 2595.0) - 1)

mel 滤波器组

def mel_filterbanks(nfilt=20, nfft=512, samplerate=16000, lowfreq=0, highfreq=None):

"""计算一个Mel-filterbank (M,F)

:param nfilt: filterbank中的滤波器数量

:param nfft: FFT size

:param samplerate: 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

"""

highfreq = highfreq or samplerate / 2

# 按梅尔均匀间隔计算点

lowmel = hz2mel(lowfreq)

highmel = hz2mel(highfreq)

melpoints = np.linspace(lowmel, highmel, nfilt + 2)

hz_points = mel2hz(melpoints) # 将mel频率再转到hz频率

# bin = samplerate/2 / NFFT/2=sample_rate/NFFT # 每个频点的频率数

# bins = hz_points/bin=hz_points*NFFT/ sample_rate # hz_points对应第几个fft频点

bin = np.floor((nfft + 1) * hz_points / samplerate)

fbank = np.zeros([nfilt, int(nfft / 2 + 1)]) # (m,f)

for i in range(0, nfilt):

for j in range(int(bin[i]), int(bin[i + 1])):

fbank[i, j] = (j - bin[i]) / (bin[i + 1] - bin[i])

for j in range(int(bin[i + 1]), int(bin[i + 2])):

fbank[i, j] = (bin[i + 2] - j) / (bin[i + 2] - bin[i + 1])

# fbank -= (np.mean(fbank, axis=0) + 1e-8)

return fbank

mel 滤波器组特征

# -*- coding:utf-8 -*-

# Author:凌逆战 | Never

# Date: 2022/5/19

"""

1、提取Mel filterBank

2、提取mel spectrum

"""

import librosa

import numpy as np

import matplotlib.pyplot as plt

import librosa.display

from matplotlib.ticker import FuncFormatter

plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签

plt.rcParams['axes.unicode_minus'] = False # 用来正常显示符号

def hz2mel(hz):

""" Hz to Mels """

return 2595 * np.log10(1 + hz / 700.0)

def mel2hz(mel):

""" Mels to HZ """

return 700 * (10 ** (mel / 2595.0) - 1)

def mel_filterbanks(nfilt=20, nfft=512, samplerate=16000, lowfreq=0, highfreq=None):

"""计算一个Mel-filterbank (M,F)

:param nfilt: filterbank中的滤波器数量

:param nfft: FFT size

:param samplerate: 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

"""

highfreq = highfreq or samplerate / 2

# 按梅尔均匀间隔计算点

lowmel = hz2mel(lowfreq)

highmel = hz2mel(highfreq)

melpoints = np.linspace(lowmel, highmel, nfilt + 2)

hz_points = mel2hz(melpoints) # 将mel频率再转到hz频率

# bin = samplerate/2 / NFFT/2=sample_rate/NFFT # 每个频点的频率数

# bins = hz_points/bin=hz_points*NFFT/ sample_rate # hz_points对应第几个fft频点

bin = np.floor((nfft + 1) * hz_points / samplerate)

fbank = np.zeros([nfilt, int(nfft / 2 + 1)]) # (m,f)

for i in range(0, nfilt):

for j in range(int(bin[i]), int(bin[i + 1])):

fbank[i, j] = (j - bin[i]) / (bin[i + 1] - bin[i])

for j in range(int(bin[i + 1]), int(bin[i + 2])):

fbank[i, j] = (bin[i + 2] - j) / (bin[i + 2] - bin[i + 1])

# fbank -= (np.mean(fbank, axis=0) + 1e-8)

return fbank

wav_path = "./p225_001.wav"

fs = 16000

NFFT = 512

win_length = 512

num_filter = 22

low_freq_mel = 0

high_freq_mel = hz2mel(fs // 2) # 求最高hz频率对应的mel频率

mel_points = np.linspace(low_freq_mel, high_freq_mel, num_filter + 2) # 在mel频率上均分成42个点

hz_points = mel2hz(mel_points) # 将mel频率再转到hz频率

print(hz_points)

# bin = sample_rate/2 / NFFT/2=sample_rate/NFFT # 每个频点的频率数

# bins = hz_points/bin=hz_points*NFFT/ sample_rate # hz_points对应第几个fft频点

bins = np.floor((NFFT + 1) * hz_points / fs)

print(bins)

# [ 0. 2. 5. 8. 12. 16. 20. 25. 31. 37. 44. 52. 61. 70.

# 81. 93. 107. 122. 138. 157. 178. 201. 227. 256.]

wav = librosa.load(wav_path, sr=fs)[0]

S = librosa.stft(wav, n_fft=NFFT, hop_length=NFFT // 2, win_length=win_length, window="hann", center=False)

mag = np.abs(S) # 幅度谱 (257, 127) librosa.magphase()

filterbanks = mel_filterbanks(nfilt=num_filter, nfft=NFFT, samplerate=fs, lowfreq=0, highfreq=fs // 2)

# ================ 画三角滤波器 ===========================

FFT_len = NFFT // 2 + 1

fs_bin = fs // 2 / (NFFT // 2) # 一个频点多少Hz

x = np.linspace(0, FFT_len, FFT_len)

plt.plot(x * fs_bin, filterbanks.T)

plt.xlim(0) # 坐标轴的范围

plt.ylim(0, 1)

plt.tight_layout()

plt.grid(linestyle='--')

plt.show()

filter_banks = np.dot(filterbanks, mag) # (M,F)*(F,T)=(M,T)

filter_banks = 20 * np.log10(filter_banks) # dB

# ================ 绘制语谱图 ==========================

# 绘制频谱图方法1

plt.imshow(filter_banks, cmap="jet", aspect='auto')

ax = plt.gca() # 获取其中某个坐标系

ax.invert_yaxis() # 将y轴反转

plt.tight_layout()

plt.show()

# 绘制频谱图方法2

plt.figure()

librosa.display.specshow(filter_banks, sr=fs, x_axis='time', y_axis='linear', cmap="jet")

plt.xlabel('时间/s', fontsize=14)

plt.ylabel('频率/kHz', fontsize=14)

plt.xticks(fontsize=14)

plt.yticks(fontsize=14)

def formatnum(x, pos):

return '$%d$' % (x / 1000)

formatter = FuncFormatter(formatnum)

plt.gca().yaxis.set_major_formatter(formatter)

plt.tight_layout()

plt.show()

画图代码

另外Librosa写好了完整的提取mel频谱和MFCC的API：

mel_spec = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, fmax=8000)

mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=40)

巴克刻度

　　巴克刻度(Bark scale)是于1961年由德国声学家Eberhard Zwicker提出的一种心理声学的尺度。它以Heinrich Barkhausen的名字命名，他提出了响度的第一个主观测量。[1]该术语的一个定义是“……等距离对应于感知上等距离的频率刻度”。高于约 500 Hz 时，此刻度或多或少等于对数频率轴。低于 500 Hz 时，Bark 标度变为越来越线性”。bark 刻度的范围是从1到24，并且它们与听觉的临界频带相对应。

频率f (Hz) 转换为 Bark：

$$\text { Bark }=13 \arctan (0.00076 f)+3.5 \arctan ((\frac{f}{7500})^{2})$$

Traunmüller, 1990 提出的新的Bark scale公式：

$$\operatorname{Bark}=\frac{26.81f}{1960+f}-0.53$$

反转：$f=\frac{1960((\operatorname{Bark}+0.53)-1)}{26.81}$

临界带宽(Hz)：$B_c=\frac{52548}{\operatorname{Bark}^2-52.56\operatorname{Bark}+690.39}$

Wang, Sekey & Gersho, 1992 提出了新的Bark scale公式：

$$\text { Bark }=6 \sinh ^{-1}(\frac{f}{600})$$

def hz2bark_1961(Hz):

return 13.0 * np.arctan(0.00076 * Hz) + 3.5 * np.arctan((Hz / 7500.0) ** 2)

def hz2bark_1990(Hz):

bark_scale = (26.81 * Hz) / (1960 + Hz) - 0.5

return bark_scale

def hz2bark_1992(Hz):

return 6 * np.arcsinh(Hz / 600)

import numpy as np

import matplotlib.pyplot as plt

from matplotlib.ticker import FuncFormatter

def hz2bark_1961(Hz):

return 13.0 * np.arctan(0.00076 * Hz) + 3.5 * np.arctan((Hz / 7500.0) ** 2)

def hz2bark_1990(Hz):

bark_scale = (26.81 * Hz) / (1960 + Hz) - 0.5

return bark_scale

def hz2bark_1992(Hz):

return 6 * np.arcsinh(Hz / 600)

if __name__ == "__main__":

fs = 16000

hz = np.linspace(0, fs // 2, fs // 2)

bark_1961 = hz2bark_1961(hz)

bark_1990 = hz2bark_1990(hz)

bark_1992 = hz2bark_1992(hz)

plt.plot(hz, bark_1961, label="bark_1961")

plt.plot(hz, bark_1990, label="bark_1990")

plt.plot(hz, bark_1992, label="bark_1992")

plt.legend() # 显示图例

plt.xlabel("Hertz scale (Hz)", fontsize=12) # x轴的名字

plt.ylabel("Bark scale", fontsize=12)

plt.xticks(fontsize=10) # x轴的刻度

plt.yticks(fontsize=10)

plt.xlim(0, fs // 2) # 坐标轴的范围

plt.ylim(0)

def formatnum(x, pos):

return '$%.1f$' % (x / 1000)

formatter = FuncFormatter(formatnum)

# plt.gca().xaxis.set_major_formatter(formatter)

# plt.gca().yaxis.set_major_formatter(formatter)

plt.grid(linestyle='--')

plt.tight_layout()

plt.show()

画图代码

Bark 滤波器组

Bark频谱

import numpy as np

import librosa

import librosa.display

import matplotlib.pyplot as plt

from matplotlib.ticker import FuncFormatter

plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签

plt.rcParams['axes.unicode_minus'] = False # 用来正常显示符号

def hz2bark(f):

""" Hz to bark频率 (Wang, Sekey & Gersho, 1992.) """

return 6. * np.arcsinh(f / 600.)

def bark2hz(fb):

""" Bark频率 to Hz """

return 600. * np.sinh(fb / 6.)

def fft2hz(fft, fs=16000, nfft=512):

""" FFT频点 to Hz """

return (fft * fs) / (nfft + 1)

def hz2fft(fb, fs=16000, nfft=512):

""" Bark频率 to FFT频点 """

return (nfft + 1) * fb / fs

def fft2bark(fft, fs=16000, nfft=512):

""" FFT频点 to Bark频率 """

return hz2bark((fft * fs) / (nfft + 1))

def bark2fft(fb, fs=16000, nfft=512):

""" Bark频率 to FFT频点 """

# bin = sample_rate/2 / nfft/2=sample_rate/nfft # 每个频点的频率数

# bins = hz_points/bin=hz_points*nfft/ sample_rate # hz_points对应第几个fft频点

return (nfft + 1) * bark2hz(fb) / fs

def Fm(fb, fc):

""" 计算一个特定的中心频率的Bark filter

:param fb: frequency in Bark.

:param fc: center frequency in Bark.

:return: 相关的Bark filter 值/幅度

"""

if fc - 2.5 <= fb <= fc - 0.5:

return 10 ** (2.5 * (fb - fc + 0.5))

elif fc - 0.5 < fb < fc + 0.5:

return 1

elif fc + 0.5 <= fb <= fc + 1.3:

return 10 ** (-2.5 * (fb - fc - 0.5))

else:

return 0

def bark_filter_banks(nfilts=20, nfft=512, fs=16000, low_freq=0, high_freq=None, scale="constant"):

""" 计算Bark-filterbanks,(B,F)

:param nfilts: 滤波器组中滤波器的数量 (Default 20)

:param nfft: FFT size.(Default is 512)

:param fs: 采样率，(Default 16000 Hz)

:param low_freq: MEL滤波器的最低带边。(Default 0 Hz)

:param high_freq: MEL滤波器的最高带边。(Default samplerate/2)

:param scale (str): 选择Max bins 幅度 "ascend"(上升)，"descend"(下降)或 "constant"(恒定)(=1)。默认是"constant"

:return:一个大小为(nfilts, nfft/2 + 1)的numpy数组，包含滤波器组。

"""

# init freqs

high_freq = high_freq or fs / 2

low_freq = low_freq or 0

# 按Bark scale 均匀间隔计算点数(点数以Bark为单位)

low_bark = hz2bark(low_freq)

high_bark = hz2bark(high_freq)

bark_points = np.linspace(low_bark, high_bark, nfilts + 4)

bins = np.floor(bark2fft(bark_points)) # Bark Scale等分布对应的 FFT bin number

# [ 0. 2. 5. 7. 10. 13. 16. 20. 24. 28. 33. 38. 44. 51.

# 59. 67. 77. 88. 101. 115. 132. 151. 172. 197. 224. 256.]

fbank = np.zeros([nfilts, nfft // 2 + 1])

# init scaler

if scale == "descendant" or scale == "constant":

c = 1

else:

c = 0

for i in range(0, nfilts): # --> B

# compute scaler

if scale == "descendant":

c -= 1 / nfilts

c = c * (c > 0) + 0 * (c < 0)

elif scale == "ascendant":

c += 1 / nfilts

c = c * (c < 1) + 1 * (c > 1)

for j in range(int(bins[i]), int(bins[i + 4])): # --> F

fc = bark_points[i+2] # 中心频率

fb = fft2bark(j) # Bark 频率

fbank[i, j] = c * Fm(fb, fc)

return np.abs(fbank)

if __name__ == "__main__":

nfilts = 22

NFFT = 512

fs = 16000

wav = librosa.load("p225_001.wav",sr=fs)[0]

S = librosa.stft(wav, n_fft=NFFT, hop_length=NFFT // 2, win_length=NFFT, window="hann", center=False)

mag = np.abs(S) # 幅度谱 (257, 127) librosa.magphase()

filterbanks = bark_filter_banks(nfilts=nfilts, nfft=NFFT, fs=fs, low_freq=0, high_freq=None, scale="constant")

# ================ 画三角滤波器 ===========================

FFT_len = NFFT // 2 + 1

fs_bin = fs // 2 / (NFFT // 2) # 一个频点多少Hz

x = np.linspace(0, FFT_len, FFT_len)

plt.plot(x * fs_bin, filterbanks.T)

# plt.xlim(0) # 坐标轴的范围

# plt.ylim(0, 1)

plt.tight_layout()

plt.grid(linestyle='--')

plt.show()

filter_banks = np.dot(filterbanks, mag) # (M,F)*(F,T)=(M,T)

filter_banks = 20 * np.log10(filter_banks) # dB

# ================ 绘制语谱图 ==========================

plt.figure()

librosa.display.specshow(filter_banks, sr=fs, x_axis='time', y_axis='linear', cmap="jet")

plt.xlabel('时间/s', fontsize=14)

plt.ylabel('频率/kHz', fontsize=14)

plt.xticks(fontsize=14)

plt.yticks(fontsize=14)

def formatnum(x, pos):

return '$%d$' % (x / 1000)

formatter = FuncFormatter(formatnum)

plt.gca().yaxis.set_major_formatter(formatter)

plt.tight_layout()

plt.show()

代码

等效矩阵带宽

　　等效矩形带宽(Equivalent Rectangular Bandwidth，ERB)是用于心理声学(研究人对声音(包括言语和音乐)的生理和心理反应的科学)的一种量度方法，它给出了一个近似于人耳听觉的对带宽的过滤方法，使用不现实但方便的简化方法将滤波器建模为矩形带通滤波器或带阻滤波器。

　　Moore 和 Glasberg在1983 年，对于中等的声强和年轻的听者，人的听觉滤波器的带宽可以通过以下的多项式方程式近似：

$$\operatorname{ERB}(f)=6.23 \cdot f^{2}+93.39 \cdot f+28.52$$

其中$f$为滤波器的中心频率(kHz)，$ERB(f)$为滤波器的带宽(Hz)。这个近似值是基于一些出版的同时掩蔽(Simultaneous masking)实验的结果。这个近似对于从0.1到6.5 kHz的范围是有效的。

　　它们也在1990年发表了另一(线性)近似：

$$\operatorname{ERB}(f)=24.7 *(4.37*10^{-3}*f+1)$$

其中$f$的单位是 Hz，$ERB(f)$的单位是 Hz。这个近似值适用于中等声级和0.1 到 10 kHz 之间的频率值。

1998发表了公式：

$$\operatorname{ERB}(f)=24.7 + \frac{f}{9.26449}$$

2002发表了公式：

\operatorname{ERB}(f)=9.265* \log(1 + \frac{f}{24.7* 9.265})

　　MATLAB的 VOICEBOX 语音处理工具箱的ERB公式：

$$\operatorname{ERBs}(f)=11.17268 \cdot \ln \left(1+\frac{46.06538 \cdot f}{f+14678.49}\right)$$

　　我看很多代码使用下面公式，但是下面公式和上面公式的

$$\operatorname{ERB}(f)=21.4 \cdot \log _{10}(1+ \frac{4.37\cdot f}{1000})$$

def hz2erb_1983(f):

""" 中心频率f(Hz) f to ERB(Hz) """

f = f / 1000.0

return (6.23 * (f ** 2)) + (93.39 * f) + 28.52

def hz2erb_1990(f):

""" 中心频率f(Hz) f to ERB(Hz) """

return 24.7 * (4.37 * f / 1000 + 1.0)

def hz2erb_1998(f):

""" 中心频率f(Hz) f to ERB(Hz)

hz2erb_1990 和 hz2erb_1990_2 的曲线几乎一模一样

M. Slaney, Auditory Toolbox, Version 2, Technical Report No: 1998-010, Internal Research Corporation, 1998

http://cobweb.ecn.purdue.edu/~malcolm/interval/1998-010/

"""

return 24.7 + (f / 9.26449)

def Hz2erb_2002(f):

""" [Hohmann2002] Equation 16 """

EarQ = 9.265 # _ERB_Q

minBW = 24.7 # minBW

return EarQ * np.log(1 + f / (minBW * EarQ))

def Hz2erb_matlab(f):

""" Convert Hz to ERB number """

n_erb = 11.17268 * np.log(1 + (46.06538 * f) / (f + 14678.49))

return n_erb

def Hz2erb_other(f):

""" Convert Hz to ERB number """

n_erb = 21.4 * np.log10(1 + 0.00437 * f)

return n_erb

其中erb_1990和erb_1998相差无几

erb202和Hz2erb_matlab和Hz2erb_other相差无几

线性滤波器组

使用ERB的线性滤波器组

# -*- coding:utf-8 -*-

# Author:凌逆战 | Never.Ling

# Date: 2022/5/28

"""

基于Josh McDermott的Matlab滤波器组代码:

https://github.com/wil-j-wil/py_bank

https://github.com/flavioeverardo/erb_bands

"""

import numpy as np

import librosa

import librosa.display

import matplotlib.pyplot as plt

from matplotlib.ticker import FuncFormatter

plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签

plt.rcParams['axes.unicode_minus'] = False # 用来正常显示符号

class EquivalentRectangularBandwidth():

def __init__(self, nfreqs, sample_rate, total_erb_bands, low_freq, max_freq):

if low_freq == None:

low_freq = 20

if max_freq == None:

max_freq = sample_rate // 2

freqs = np.linspace(0, max_freq, nfreqs) # 每个STFT频点对应多少Hz

self.EarQ = 9.265 # _ERB_Q

self.minBW = 24.7 # minBW

# 在ERB刻度上建立均匀间隔

erb_low = self.freq2erb(low_freq) # 最低截止频率

erb_high = self.freq2erb(max_freq) # 最高截止频率

# 在ERB频率上均分为(total_erb_bands + )2个频带

erb_lims = np.linspace(erb_low, erb_high, total_erb_bands + 2)

cutoffs = self.erb2freq(erb_lims) # 将 ERB频率再转到 hz频率, 在线性频率Hz上找到ERB截止频率对应的频率

# self.nfreqs F

# self.freqs # 每个STFT频点对应多少Hz

self.filters = self.get_bands(total_erb_bands, nfreqs, freqs, cutoffs)

def freq2erb(self, frequency):

""" [Hohmann2002] Equation 16"""

return self.EarQ * np.log(1 + frequency / (self.minBW * self.EarQ))

def erb2freq(self, erb):

""" [Hohmann2002] Equation 17"""

return (np.exp(erb / self.EarQ) - 1) * self.minBW * self.EarQ

def get_bands(self, total_erb_bands, nfreqs, freqs, cutoffs):

"""

获取erb bands、索引、带宽和滤波器形状

:param erb_bands_num: ERB 频带数

:param nfreqs: 频点数 F

:param freqs: 每个STFT频点对应多少Hz

:param cutoffs: 中心频率 Hz

:param erb_points: ERB频带界限列表

:return:

"""

cos_filts = np.zeros([nfreqs, total_erb_bands]) # (F, ERB)

for i in range(total_erb_bands):

lower_cutoff = cutoffs[i] # 上限截止频率 Hz

higher_cutoff = cutoffs[i + 2] # 下限截止频率 Hz, 相邻filters重叠50%

lower_index = np.min(np.where(freqs > lower_cutoff)) # 下限截止频率对应的Hz索引 Hz。np.where 返回满足条件的索引

higher_index = np.max(np.where(freqs < higher_cutoff)) # 上限截止频率对应的Hz索引

avg = (self.freq2erb(lower_cutoff) + self.freq2erb(higher_cutoff)) / 2

rnge = self.freq2erb(higher_cutoff) - self.freq2erb(lower_cutoff)

cos_filts[lower_index:higher_index + 1, i] = np.cos(

(self.freq2erb(freqs[lower_index:higher_index + 1]) - avg) / rnge * np.pi) # 减均值，除方差

# 加入低通和高通，得到完美的重构

filters = np.zeros([nfreqs, total_erb_bands + 2]) # (F, ERB)

filters[:, 1:total_erb_bands + 1] = cos_filts

# 低通滤波器上升到第一个余cos filter的峰值

higher_index = np.max(np.where(freqs < cutoffs[1])) # 上限截止频率对应的Hz索引

filters[:higher_index + 1, 0] = np.sqrt(1 - np.power(filters[:higher_index + 1, 1], 2))

# 高通滤波器下降到最后一个cos filter的峰值

lower_index = np.min(np.where(freqs > cutoffs[total_erb_bands]))

filters[lower_index:nfreqs, total_erb_bands + 1] = np.sqrt(

1 - np.power(filters[lower_index:nfreqs, total_erb_bands], 2))

return cos_filts

if __name__ == "__main__":

fs = 16000

NFFT = 512 # 信号长度

ERB_num = 20

low_lim = 20 # 最低滤波器中心频率

high_lim = fs / 2 # 最高滤波器中心频率

freq_num = NFFT // 2 + 1

fs_bin = fs // 2 / (NFFT // 2) # 一个频点多少Hz

x = np.linspace(0, freq_num, freq_num)

# ================ 画三角滤波器 ===========================

ERB = EquivalentRectangularBandwidth(freq_num, fs, ERB_num, low_lim, high_lim)

filterbanks = ERB.filters.T # (257, 20)

plt.plot(x * fs_bin, filterbanks.T)

# plt.xlim(0) # 坐标轴的范围

# plt.ylim(0, 1)

plt.tight_layout()

plt.grid(linestyle='--')

plt.show()

# ================ 绘制语谱图 ==========================

wav = librosa.load("p225_001.wav", sr=fs)[0]

S = librosa.stft(wav, n_fft=NFFT, hop_length=NFFT // 2, win_length=NFFT, window="hann", center=False)

mag = np.abs(S) # 幅度谱 (257, 127) librosa.magphase()

filter_banks = np.dot(filterbanks, mag) # (M,F)*(F,T)=(M,T)

filter_banks = 20 * np.log10(filter_banks) # dB

plt.figure()

librosa.display.specshow(filter_banks, sr=fs, x_axis='time', y_axis='linear', cmap="jet")

plt.xlabel('时间/s', fontsize=14)

plt.ylabel('频率/kHz', fontsize=14)

plt.xticks(fontsize=14)

plt.yticks(fontsize=14)

def formatnum(x, pos):

return '$%d$' % (x / 1000)

formatter = FuncFormatter(formatnum)

plt.gca().yaxis.set_major_formatter(formatter)

plt.tight_layout()

plt.show()

View Code

Gammatone 滤波器组

　　外界语音信号进入耳蜗的基底膜后，将依据频率进行分解并产生行波震动，从而刺激听觉感受细胞。GammaTone 滤波器是一组用来模拟耳蜗频率分解特点的滤波器模型，由脉冲响应描述的线性滤波器，脉冲响应是gamma 分布和正弦(sin)音调的乘积。它是听觉系统中一种广泛使用的听觉滤波器模型。

历史

　　一般认为外周听觉系统的频率分析方式可以通过一组带通滤波器来进行一定程度的模拟，人们为此也提出了各种各样的滤波器组，如 roex 滤波器(Patterson and Moore 1986)。

在神经科学上有一种叫做反向相关性 “reverse correlation”(de Boer and Kuyper 1968)的计算方式，通过计算初级听觉神经纤维对于白噪声刺激的响应以及相关程度，即听觉神经元发放动作电位前的平均叠加信号，从而直接从生理状态上估计听觉滤波器的形状。这个滤波器是在外周听觉神经发放动作电位前生效的，因此得名为“revcor function”，可以作为一定限度下对外周听觉滤波器冲激响应的估计，也就是耳蜗等对音频信号的前置带通滤波。

　　1972年Johannesma提出了 gammatone 滤波器用来逼近recvor function：

$$时域表达式：g(t)=a t^{n-1} e^{-2 \pi b t} \cos (2 \pi f_c t+\phi_0)$$

其中$f_c(Hz)$是中心频率(center frequency)，$\phi_0$是初始相位(phase)，$a$是幅度(amplitude)，$n$是滤波器的阶数(order)，越大则偏度越低，滤波器越“瘦高”，$b(Hz)$是滤波器的3dB 带宽(bandwidth)，$t(s)$是时间。

这个时域脉冲响应是一个正弦曲线(pure tone)，其幅度包络是一个缩放的gamma分布函数。

我们可以通过时域表达式生成一组gammatone滤波器组和 gammatone滤波器组特征。

# -*- coding:utf-8 -*-

# Author:凌逆战 | Never.Ling

# Date: 2022/5/24

"""

时域滤波器组 FFT 转频域滤波器组与语音频谱相乘

参考：https://github.com/TAriasVergara/Acoustic_features

"""

import librosa

import librosa.display

import numpy as np

from scipy.fftpack import dct

import matplotlib.pyplot as plt

from matplotlib.ticker import FuncFormatter

plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签

plt.rcParams['axes.unicode_minus'] = False # 用来正常显示符号

def erb_space(low_freq=50, high_freq=8000, n=64):

""" 计算中心频率(ERB scale)

:param min_freq: 中心频率域的最小频率。

:param max_freq: 中心频率域的最大频率。

:param nfilts: 滤波器的个数，即等于计算中心频率的个数。

:return: 一组中心频率

"""

ear_q = 9.26449

min_bw = 24.7

cf_array = -(ear_q * min_bw) + np.exp(

np.linspace(1, n, n) * (-np.log(high_freq + ear_q * min_bw) + np.log(low_freq + ear_q * min_bw)) / n) \

* (high_freq + ear_q * min_bw)

return cf_array

def gammatone_impulse_response(samplerate_hz, length_in_samples, center_freq_hz, p):

""" gammatone滤波器的时域公式

:param samplerate_hz: 采样率

:param length_in_samples: 信号长度

:param center_freq_hz: 中心频率

:param p: 滤波器阶数

:return: gammatone 脉冲响应

"""

# 生成一个gammatone filter (1990 Glasberg&Moore parametrized)

erb = 24.7 + (center_freq_hz / 9.26449) # equivalent rectangular bandwidth.

# 中心频率

an = (np.pi * np.math.factorial(2 * p - 2) * np.power(2, float(-(2 * p - 2)))) / np.square(np.math.factorial(p - 1))

b = erb / an # 带宽

a = 1 # 幅度(amplitude). 这在后面的归一化过程中有所不同。

t = np.linspace(1. / samplerate_hz, length_in_samples / samplerate_hz, length_in_samples)

gammatone_ir = a * np.power(t, p - 1) * np.exp(-2 * np.pi * b * t) * np.cos(2 * np.pi * center_freq_hz * t)

return gammatone_ir

def generate_filterbank(fs, fmax, L, N, p=4):

"""

L: 在样本中测量的信号的大小

N: 滤波器数量

p: Gammatone脉冲响应的阶数

"""

# 中心频率

if fs == 8000:

fmax = 4000

center_freqs = erb_space(50, fmax, N) # 中心频率列表

center_freqs = np.flip(center_freqs) # 反转数组

n_center_freqs = len(center_freqs) # 中心频率的数量

filterbank = np.zeros((N, L))

# 为每个中心频率生成滤波器

for i in range(n_center_freqs):

# aa = gammatone_impulse_response(fs, L, center_freqs[i], p)

filterbank[i, :] = gammatone_impulse_response(fs, L, center_freqs[i], p)

return filterbank

def gfcc(cochleagram, numcep=13):

feat = dct(cochleagram, type=2, axis=1, norm='ortho')[:, :numcep]

# feat-= (np.mean(feat, axis=0) + 1e-8)#Cepstral mean substration

return feat

def cochleagram(sig_spec, filterbank, nfft):

"""

:param sig_spec: 语音频谱

:param filterbank: 时域滤波器组

:param nfft: fft_size

:return:

"""

filterbank = powerspec(filterbank, nfft) # 时域滤波器组经过FFT变换

filterbank /= np.max(filterbank, axis=-1)[:, None] # Normalize filters

cochlea_spec = np.dot(sig_spec, filterbank.T) # 矩阵相乘

cochlea_spec = np.where(cochlea_spec == 0.0, np.finfo(float).eps, cochlea_spec) # 把0变成一个很小的数

# cochlea_spec= np.log(cochlea_spec)-np.mean(np.log(cochlea_spec),axis=0)

cochlea_spec = np.log(cochlea_spec)

return cochlea_spec, filterbank

def powerspec(X, nfft):

# Fourier transform

# Y = np.fft.rfft(X, n=n_padded)

Y = np.fft.fft(X, n=nfft)

Y = np.absolute(Y)

# non-redundant part

m = int(nfft / 2) + 1

Y = Y[:, :m]

return np.abs(Y) ** 2

if __name__ == "__main__":

nfilts = 22

NFFT = 512

fs = 16000

Order = 4

FFT_len = NFFT // 2 + 1

fs_bin = fs // 2 / (NFFT // 2) # 一个频点多少Hz

x = np.linspace(0, FFT_len, FFT_len)

# ================ 画三角滤波器 ===========================

# gammatone_impulse_response = gammatone_impulse_response(fs/2, 512, 200, Order) # gammatone冲击响应

generate_filterbank = generate_filterbank(fs, fs // 2, FFT_len, nfilts, Order)

filterbanks = powerspec(generate_filterbank, NFFT) # 时域滤波器组经过FFT变换

filterbanks /= np.max(filterbanks, axis=-1)[:, None] # Normalize filters

print(generate_filterbank.shape) # (22, 257)

# plt.plot(filterbanks.T)

plt.plot(x * fs_bin, filterbanks.T)

# plt.xlim(0) # 坐标轴的范围

# plt.ylim(0, 1)

plt.tight_layout()

plt.grid(linestyle='--')

plt.show()

# ================ 绘制语谱图 ==========================

wav = librosa.load("p225_001.wav", sr=fs)[0]

S = librosa.stft(wav, n_fft=NFFT, hop_length=NFFT // 2, win_length=NFFT, window="hann", center=False)

mag = np.abs(S) # 幅度谱 (257, 127) librosa.magphase()

filter_banks = np.dot(filterbanks, mag) # (M,F)*(F,T)=(M,T)

filter_banks = 20 * np.log10(filter_banks) # dB

plt.figure()

librosa.display.specshow(filter_banks, sr=fs, x_axis='time', y_axis='linear', cmap="jet")

plt.xlabel('时间/s', fontsize=14)

plt.ylabel('频率/kHz', fontsize=14)

plt.xticks(fontsize=14)

plt.yticks(fontsize=14)

def formatnum(x, pos):

return '$%d$' % (x / 1000)

formatter = FuncFormatter(formatnum)

plt.gca().yaxis.set_major_formatter(formatter)

plt.tight_layout()

plt.show()

View Code

　　可以看到低频段分得很细，高频段分得很粗，和人耳听觉特性较为符合。

$$频域表达式：\begin{aligned} H(f)=& a[R(f) \otimes S(f)] \\ =& \frac{a}{2}(n-1) !(2 \pi b)^{-n}\left\{e^{i \phi_0}\left[1+\frac{i(f-f_c)}{b} \right]^{-n}+e^{-i \phi_0}\left[1+\frac{i(f+f_c)}{b}\right]^{-n}\right\} \end{aligned}$$

频率表达式中$R(f)$是指数+阶跃函数的傅里叶变换，阶跃函数用来区别 t>0 和 t<0。$S(f)$是频率为$f_0$的余弦的傅里叶变换。可以看到是一个中心频率在$f_c$、在两侧按照e指数衰减的滤波器。通过上述表达式可以生成一组滤波器，求Gammatone滤波器组特征只需要将Gammatone滤波器组与语音幅度谱相乘即可得到Gammatone滤波器组特征。

# -*- coding:utf-8 -*-

# Author:凌逆战 | Never

# Date: 2022/5/24

"""

Gammatone-filter-banks implementation

based on https://github.com/mcusi/gammatonegram/

"""

import librosa

import librosa.display

import numpy as np

from matplotlib import pyplot as plt

from matplotlib.ticker import FuncFormatter

plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签

plt.rcParams['axes.unicode_minus'] = False # 用来正常显示符号

# Slaney's ERB Filter constants

EarQ = 9.26449

minBW = 24.7

def generate_center_frequencies(min_freq, max_freq, filter_nums):

""" 计算中心频率(ERB scale)

:param min_freq: 中心频率域的最小频率。

:param max_freq: 中心频率域的最大频率。

:param filter_nums: 滤波器的个数，即等于计算中心频率的个数。

:return: 一组中心频率

"""

# init vars

n = np.linspace(1, filter_nums, filter_nums)

c = EarQ * minBW

# 计算中心频率

cfreqs = (max_freq + c) * np.exp((n / filter_nums) * np.log(

(min_freq + c) / (max_freq + c))) - c

return cfreqs

def compute_gain(fcs, B, wT, T):

""" 为了阶数计算增益和矩阵化计算

:param fcs: 中心频率

:param B: 滤波器的带宽

:param wT: 对应于用于频域计算的 w * T = 2 * pi * freq * T

:param T: 周期(单位秒s)，1/fs

:return:

Gain: 表示filter gains 的2d numpy数组

A: 用于最终计算的二维数组

"""

# 为了简化预先计算

K = np.exp(B * T)

Cos = np.cos(2 * fcs * np.pi * T)

Sin = np.sin(2 * fcs * np.pi * T)

Smax = np.sqrt(3 + 2 ** (3 / 2))

Smin = np.sqrt(3 - 2 ** (3 / 2))

# 定义A矩阵的行

A11 = (Cos + Smax * Sin) / K

A12 = (Cos - Smax * Sin) / K

A13 = (Cos + Smin * Sin) / K

A14 = (Cos - Smin * Sin) / K

# 计算增益 (vectorized)

A = np.array([A11, A12, A13, A14])

Kj = np.exp(1j * wT)

Kjmat = np.array([Kj, Kj, Kj, Kj]).T

G = 2 * T * Kjmat * (A.T - Kjmat)

Coe = -2 / K ** 2 - 2 * Kj ** 2 + 2 * (1 + Kj ** 2) / K

Gain = np.abs(G[:, 0] * G[:, 1] * G[:, 2] * G[:, 3] * Coe ** -4)

return A, Gain

def gammatone_filter_banks(nfilts=22, nfft=512, fs=16000, low_freq=None, high_freq=None, scale="contsant", order=4):

""" 计算Gammatone-filterbanks, (G,F)

:param nfilts: filterbank中滤波器的数量 (Default 22)

:param nfft: FFT size (Default is 512)

:param fs: 采样率 (Default 16000 Hz)

:param low_freq: 最低频带 (Default 0 Hz)

:param high_freq: 最高频带 (Default samplerate/2)

:param scale: 选择Max bins 幅度 "ascend"(上升)，"descend"(下降)或 "constant"(恒定)(=1)。默认是"constant"

:param order: 滤波器阶数

:return: 一个大小为(nfilts, nfft/2 + 1)的numpy数组，包含滤波器组。

"""

# init freqs

high_freq = high_freq or fs / 2

low_freq = low_freq or 0

# define custom difference func

def Dif(u, a):

return u - a.reshape(nfilts, 1)

# init vars

fbank = np.zeros([nfilts, nfft])

width = 1.0

maxlen = nfft // 2 + 1

T = 1 / fs

n = 4

u = np.exp(1j * 2 * np.pi * np.array(range(nfft // 2 + 1)) / nfft)

idx = range(nfft // 2 + 1)

fcs = generate_center_frequencies(low_freq, high_freq, nfilts) # 计算中心频率，转换到ERB scale

ERB = width * ((fcs / EarQ) ** order + minBW ** order) ** (1 / order) # 计算带宽

B = 1.019 * 2 * np.pi * ERB

# compute input vars

wT = 2 * fcs * np.pi * T

pole = np.exp(1j * wT) / np.exp(B * T)

# compute gain and A matrix

A, Gain = compute_gain(fcs, B, wT, T)

# compute fbank

fbank[:, idx] = (

(T ** 4 / Gain.reshape(nfilts, 1)) *

np.abs(Dif(u, A[0]) * Dif(u, A[1]) * Dif(u, A[2]) * Dif(u, A[3])) *

np.abs(Dif(u, pole) * Dif(u, pole.conj())) ** (-n))

# 确保所有filters的最大值为1.0

try:

fbs = np.array([f / np.max(f) for f in fbank[:, range(maxlen)]])

except BaseException:

fbs = fbank[:, idx]

# compute scaler

if scale == "ascendant":

c = [

]

for i in range(1, nfilts):

x = c[i - 1] + 1 / nfilts

c.append(x * (x < 1) + 1 * (x > 1))

elif scale == "descendant":

c = [

]

for i in range(1, nfilts):

x = c[i - 1] - 1 / nfilts

c.append(x * (x > 0) + 0 * (x < 0))

else:

c = [1 for i in range(nfilts)]

# apply scaler

c = np.array(c).reshape(nfilts, 1)

fbs = c * np.abs(fbs)

return fbs

if __name__ == "__main__":

nfilts = 22

NFFT = 512

fs = 16000

FFT_len = NFFT // 2 + 1

fs_bin = fs // 2 / (NFFT // 2) # 一个频点多少Hz

x = np.linspace(0, FFT_len, FFT_len)

# ================ 画三角滤波器 ===========================

filterbanks = gammatone_filter_banks(nfilts=22, nfft=512, fs=16000,

low_freq=None, high_freq=None,

scale="contsant", order=4)

print(filterbanks.shape) # (22, 257)

plt.plot(x * fs_bin, filterbanks.T)

# plt.xlim(0) # 坐标轴的范围

# plt.ylim(0, 1)

plt.tight_layout()

plt.grid(linestyle='--')

plt.show()

# ================ 绘制语谱图 ==========================

wav = librosa.load("p225_001.wav", sr=fs)[0]

S = librosa.stft(wav, n_fft=NFFT, hop_length=NFFT // 2, win_length=NFFT, window="hann", center=False)

mag = np.abs(S) # 幅度谱 (257, 127) librosa.magphase()

filter_banks = np.dot(filterbanks, mag) # (M,F)*(F,T)=(M,T)

filter_banks = 20 * np.log10(filter_banks) # dB

plt.figure()

librosa.display.specshow(filter_banks, sr=fs, x_axis='time', y_axis='linear', cmap="jet")

plt.xlabel('时间/s', fontsize=14)

plt.ylabel('频率/kHz', fontsize=14)

plt.xticks(fontsize=14)

plt.yticks(fontsize=14)

def formatnum(x, pos):

return '$%d$' % (x / 1000)

formatter = FuncFormatter(formatnum)

plt.gca().yaxis.set_major_formatter(formatter)

plt.tight_layout()

plt.show()

View Code

　　1988年Holdsworth 等人进一步阐明了GTF的各种特性，而且提供了一个数字IIR滤波器设计方案。这个技术使得GTF能够比FIR更加容易且高效地实现，为后续出现一些重要的实际应用做了铺垫。听觉滤波的gammatone模型的变化和改进包括复数gammatone滤波器、gammachirp滤波器、全极点(all-pole)和一零(one-zero) gammatone滤波器、双边(two-sided)gammatone滤波器和滤波器级联(filter-cascade)模型，以及各种level相关和这些的动态非线性版本。

参考

【博客】Auditory scales of frequency representation

【百度百科】心理声学

【维基百科】Bark scale

【维基百科】Mel scale

【维基百科】Equivalent rectangular bandwidth

【维基百科】Gammatone filter(包含了C \ C++ \ mathematica \ matlab的代码实现)

【博客】Equivalent Rectangular Bandwidth

【CSDN】GammaTone 滤波器详解

【python库】PyFilterbank

【代码】Brookes, Mike (22 December 2012). "frq2erb". VOICEBOX: Speech Processing Toolbox for MATLAB. Department of Electrical & Electronic Engineering, Imperial College, UK. Retrieved 20 January 2013.

【代码】Brookes, Mike (22 December 2012). "erb2frq". VOICEBOX: Speech Processing Toolbox for MATLAB. Department of Electrical & Electronic Engineering, Imperial College, UK. Retrieved 20 January 2013.

【论文】Smith, Julius O.; Abel, Jonathan S. (10 May 2007). "Equivalent Rectangular Bandwidth". Bark and ERB Bilinear Transforms. Center for Computer Research in Music and Acoustics (CCRMA), Stanford University, USA. Retrieved 20 January 2013.

关注博主即可阅读全文

优惠劵

凌逆战

关注

踩

觉得还不错?

一键收藏

打赏

知道了

声学感知刻度(mel scale、Bark scale、ERB)与声学特征提取(MFCC、BFCC、GFCC)

本文地址：https://www.cnblogs.com/LXP-Never/p/16011229.html (引用请注明出处)本文代码：https://github.com/LXP-Never/perception_scale作者：凌逆战 | Never.Ling梅尔刻度　　梅尔刻度(Mel scale)是一种由听众判断不同频率音高(pitch)彼此相等的感知刻度，表示人耳对等距......

复制链接

扫一扫

[语音识别]声学特征提取

小哲的博客

04-08

4685

声学特征提取语音识别：声学特征提取1. 预加重2. 分帧3. 加窗4. 离散傅里叶变换（DFT）5. 语谱图6. 梅尔刻度（Mel Scale）7. 梅尔滤波器组（Mel Filter Bank）8. FBANK特征9. MFCC特征10. 差分11. CQCC特征总结

语音识别：声学特征提取

常用的声学特征有FBANK、MFCC、PLP等， MFCC特征各纬度之间具有较弱的相关性，适合GMM的训练，FBANK相比MFCC保留了更原始的声学特征，多用于DNN的训练。

1. 预加重

语音中有频谱倾斜现象，即

GFCC和MFCC特征提取（python代码）

04-20

提取语音的GFCC特征，不需要搭建环境，可以直接运行，希望大家支持一下。如果下载后不可以使用，可以csdn联系我

3 条评论

您还未登录，请先

后发表或查看评论

【无标题】

梅逊雪——记录科学研究

08-10

https://blog.51cto.com/u_15069443/2922889

音频筑基：巴克谱和梅尔谱辨析

来知晓的博客

01-03

663

在音频信号处理中，巴克谱和梅尔谱是我们经常遇到的概念，也是语音处理中常用到的频域特征，这里谈谈自己对它们的理解。

GFCC

weixin_39011425的博客

11-14

5059

声学特征: GFCC

1. introduction

ASA

CASA

2. Auditory Features

input（signal） -> STFT -> Gammatone filters -> downsampling（改变采样频率到10KHz） -> loudness-compressed (减少 magnitude) -> output（TF deco...

GFCC的matla实现

10-17

根据GFCC的一般实现流程，利用matlab实现算法。此程序可以有效的对音频信号处理。

声学特征提取普及笔记

weixin_38858860的博客

01-08

865

声学特征预处理：预加重、分帧、加窗；声学特征提取：STFT、FBank、语谱图、MFCC

声音信号分离

深入理解梅尔刻度、梅尔滤波器组和梅尔时频谱图

DEDSEC_Roger的博客

12-10

3726

读取一段音频，使用短时傅里叶变换，得到普通的时频谱图，然后绘制梅尔滤波器组，值得注意的是，librosa的梅尔滤波器组函数还带有权重归一化功能，即对一个三角形滤波器的每个权重，都除以该三角形的面积，如果不希望进行该归一化，设置参数。从公式可见，对数部分可以以自然对数为底数，也可以以10为底数，不同的底数对应不同的系数，要确定当前的系数，只需要代入(1000Hz, 1000mel)即可。其中，m是当前滤波器的序号，表征了当前的时间段，k是当前频率的序号，表征了当前正在对哪一频率的。

临界频带和听觉滤波器

weixin_47546087的博客

08-31

1946

设计滤波器组时，经常用ERB尺度作为频带划分标准，中心频率取ERB尺度的1到40整数值等间隔。右图是基于多个测量数据拟合的ERB带宽和线性频率的关系曲线，以及与传统Bark临界频带（点线）的对比，频率单位kHz，适用于中心频率100Hz—10kHz之间的情况。之后临界带宽随着频率增加而增大，与0.2f虚线作对比，开始临界带宽增大慢一些，3kHz以后临界带宽增大变得更快。1、线性频率—听觉心理尺度的转换：将频谱展示的线性频率横轴，转换为Bark尺度值、ERB尺度值、倍频程序号、1/3倍频程序号；......

Gammatone 滤波器的 Python 代码实现

GFCC特征参数改进建议 - CSDN文库

特征参数改进建议 - CSDN文库

首页GFCC特征参数改进建议 GFCC特征参数改进建议时间: 2023-09-15 17:17:08 浏览: 152 GFCC（Gammatone Frequency Cepstral Coefficients）特征参数是一种用于音频信号处理的特征提取方法。在进行GFCC特征参数改进时，可以考虑以下几点建议：

1. 增加滤波器组数：当前GFCC特征参数通常使用40个滤波器组，可以尝试增加滤波器组数，以提高对音频信号的分辨率和细节处理能力。

2. 调整滤波器参数：GFCC特征参数中，滤波器的参数（如中心频率、带宽等）对特征提取效果有很大影响，可以通过调整滤波器的参数，来优化GFCC特征参数的表现。

3. 引入局部归一化：由于音频信号具有时域和频域的特性，可以考虑在GFCC特征参数中引入局部归一化，以更好地处理信号的时频特性，并提高特征的稳定性。

4. 尝试其他特征参数方法：除了GFCC特征参数，还有很多其他的特征参数提取方法，如MFCC、PLP等，可以尝试这些方法，以获得更好的特征表现。相关问题提取gfcc特征代码 GFCC（gammatone frequency cepstral coefficients）特征是一种音频特征提取方法，是对音频信号进行频谱处理后进行MFCC（Mel频率倒谱系数）的改进。提取GFCC特征的代码可分为以下几个步骤：

1. 预处理：读取音频信号，进行预处理操作，例如归一化、去除静音段等。

2. 帧化：将预处理后的音频信号分成帧，每帧通常选取20-40毫秒的时间长度，帧与帧之间有一定的重叠。

3. 加窗：对每一帧的音频信号应用窗函数（如汉明窗），以减少由帧分割引起的频谱泄漏。

4. 快速傅里叶变换（FFT）：对窗口化的音频信号进行FFT变换，将信号转换为频域表示。

5. 滤波器组设gfcc特征提取matlab 相关推荐 GFCC和MFCC特征提取（python代码）提取语音的GFCC特征，不需要搭建环境，可以直接运行，希望大家支持一下。如果下载后不可以使用，可以csdn联系我 MATLAB提取MFCC、GFCC、LPCC等特征，使用随机森林分类 MATLAB首先对语音进行不同的非线性自适应时频分析的去噪，然后提取MFCC、GFCC、LPCC等特征，最后通过随机森林，对音标进行分类注1：音频文件数据集；注2：一行代码自动添加文件和子文件到路径； GFCC的matla实现根据GFCC的一般实现流程，利用matlab实现算法。此程序可以有效的对音频信号处理。 matlab的gfcc特征提取要在MATLAB中提取GFCC（Gammatone Frequency Cepstral Coefficients）特征，可以使用以下代码： matlab % 首先对语音进行非线性自适应时频分析的去噪 noisySignal = audioIn; % 输入的语音信号 cleanSignal = ... 语音特征GFCC和MFCC融合的建议将语音特征GFCC和MFCC融合可以提高语音信号的识别性能，具体建议如下： 1. 将GFCC和MFCC的特征向量拼接起来作为一个新的特征向量，然后输入到分类器中进行训练和测试。 2. 在GFCC和MFCC的特征向量上分别训练不同的... 语音特征MFCC与GFCC融合的建议 MFCC和GFCC都是常用的语音特征提取方法，它们在不同的应用场景下表现出不同的优势。融合MFCC和GFCC可以综合利用它们的优势，提高语音识别的准确率。一种常用的方法是将MFCC和GFCC的特征向量拼接起来，形成一个更长... GFCC python GFCC (Gammatone Frequency Cepstral Coefficients) 是一种用于音频信号处理的特征提取方法。它是从Gammatone滤波器组输出中计算的，用于捕捉音频信号中的频率特征。在Python中，你可以使用Librosa库来计算GFCC特征... mfcc、bfcc、gfcc mfcc、bfcc和gfcc都是基于人耳听觉模型的信号特征提取方法。它们的主要作用是将语音信号转换为具有可分辨语音信息的特征向量。 MFCC（Mel Frequency Cepstral Coefficients）是最早被广泛研究和应用的一种，它主要... 随机森林特征提取 MATLAB 然后，使用MFCC、GFCC、LPCC等特征提取方法从语音信号中提取特征。最后，利用随机森林算法对这些特征进行分类。可以使用一行代码将音频文件和子文件添加到MATLAB的路径中，以便在处理过程中能够方便地读取和处理... GFCC和MFCC特征提取附python代码+仿真结果和运行方法.zip 1.版本：matlab2014/2019a/2021a，内含运行结果，不会运行可私信 2.领域：智能优化算法、神经网络预测、信号处理、元胞自动机、图像处理、路径规划、无人机等多种领域的Matlab仿真，更多内容可点击博主头像 ... gfcc-speech-kaldi gfcc-speech-kaldi 如果您使用此代码或部分代码，请引用我们！ Puneet Bawa，Virender Kadyan，“在不匹配条件下用于自动旁遮普识别系统的噪声强大的域内儿童语音增强功能” doi：： GFCC.zip_GFCC_GFCC提取_gfcc python_python GFCC_python实现gfcc 提取语音的GFCC，获得参数，对音频信号进行处理语音信号处理- MFCC特征提取 (1)掌握MFCC原理； (2)学会使用MATLAB编程进行MFCC特征提取。 GFCC.zip_GFCC_XV7U_breezetep_features extraction_scientistii6 4/5000 Acoustics features extraction. GFCC pandas_redshift-1.0.2.tar.gz Python库是一组预先编写的代码模块，旨在帮助开发者实现特定的编程任务，无需从零开始编写代码。这些库可以包括各种功能，如数学运算、文件操作、数据分析和网络编程等。Python社区提供了大量的第三方库，如NumPy、Pandas和Requests，极大地丰富了Python的应用领域，从数据科学到Web开发。Python库的丰富性是Python成为最受欢迎的编程语言之一的关键原因之一。这些库不仅为初学者提供了快速入门的途径，而且为经验丰富的开发者提供了强大的工具，以高效率、高质量地完成复杂任务。例如，Matplotlib和Seaborn库在数据可视化领域内非常受欢迎，它们提供了广泛的工具和技术，可以创建高度定制化的图表和图形，帮助数据科学家和分析师在数据探索和结果展示中更有效地传达信息。

实战大数据扩展视频Part3.zip.009 实战大数据扩展视频Part3.zip.009 pandas_plink-2.2.7-cp38-cp38-manylinux1_x86_64.whl Python库是一组预先编写的代码模块，旨在帮助开发者实现特定的编程任务，无需从零开始编写代码。这些库可以包括各种功能，如数学运算、文件操作、数据分析和网络编程等。Python社区提供了大量的第三方库，如NumPy、Pandas和Requests，极大地丰富了Python的应用领域，从数据科学到Web开发。Python库的丰富性是Python成为最受欢迎的编程语言之一的关键原因之一。这些库不仅为初学者提供了快速入门的途径，而且为经验丰富的开发者提供了强大的工具，以高效率、高质量地完成复杂任务。例如，Matplotlib和Seaborn库在数据可视化领域内非常受欢迎，它们提供了广泛的工具和技术，可以创建高度定制化的图表和图形，帮助数据科学家和分析师在数据探索和结果展示中更有效地传达信息。

CSDN会员开通CSDN年卡参与万元壕礼抽奖海量 VIP免费资源千本正版电子书商城会员专享价千门课程&专栏全年可省5,000元立即开通全年可省5,000元立即开通最新推荐 pandas_redshift-1.0.2.tar.gz Python库是一组预先编写的代码模块，旨在帮助开发者实现特定的编程任务，无需从零开始编写代码。这些库可以包括各种功能，如数学运算、文件操作、数据分析和网络编程等。Python社区提供了大量的第三方库，如NumPy、Pandas和Requests，极大地丰富了Python的应用领域，从数据科学到Web开发。Python库的丰富性是Python成为最受欢迎的编程语言之一的关键原因之一。这些库不仅为初学者提供了快速入门的途径，而且为经验丰富的开发者提供了强大的工具，以高效率、高质量地完成复杂任务。例如，Matplotlib和Seaborn库在数据可视化领域内非常受欢迎，它们提供了广泛的工具和技术，可以创建高度定制化的图表和图形，帮助数据科学家和分析师在数据探索和结果展示中更有效地传达信息。

红色响应式生鲜肉类电商bootstrap模板.zip 【技术分析】

HTML，是一种制作万维网页面的标准语言，它消除了不同计算机之间信息交流的障碍；

CSS，可以帮助把网页外观做得更加美观；

JavaScript，是一种轻量级的解释型编程语言；

jQuery，使用户能更方便地处理HTML documents、events、实现动画效果，并且方便地为网站提供AJAX交互；

Bootstrap 是快速开发 Web 应用程序的前端工具包。它是一个 CSS，HTML 和 JS 的集合，它使用了最新的浏览器技术，给你的 Web 开发提供了时尚的版式；

AJAX，创建交互式网页应用的网页开发技术。

【设计思路】

用户友好性：界面简洁直观，易于操作，减少用户的学习成本。

模块化设计：将系统功能模块化，每个模块负责一类功能，方便扩展和维护。

Responsiveness：后台管理系统应当是响应式设计，能够适配不同设备屏幕大小，包括电脑、平板和手机等。

权限控制：根据用户角色设定不同的权限，确保用户只能访问其权限范围内的功能。

数据安全：对用户数据进行加密存储、访问控制等措施，保护用户隐私和系统安全。

日志功能：记录关键操作日志，保留操作痕迹。 Android移动开发期末大作业.zip Android移动开发期末大作业.zip 电力系统自动化(4频率与有功调节).pptx 电力系统自动化是指利用现代控制和通信技术，对电力系统中的各种设备和分系统进行智能监控、保护和自动调节的过程。电力系统自动化旨在提高系统的运行效率、可靠性和安全性，同时实现对系统的精准控制和管理。

在电力系统自动化中，频率与有功功率调节是一个重要的方面。频率是指电力系统中电压和功率波动的频率，是电能质量的重要指标之一。频率偏移是电力系统有功功率不平衡的结果，需要通过调频来实现功率平衡。频率调节可以在电源侧和负荷侧进行，分别通过调节机组的有功输出和切除或增加负荷来实现频率的稳定。

频率调节的意义和特点在于，频率调节可以影响整个电力系统的运行状态和电能质量，是一个系统级的自动控制过程。频率调节不仅需要考虑系统的稳定性和安全性，还需要与经济运行密切相关。因此，在电力系统自动化中，频率调节是一个复杂而重要的问题，需要综合考虑系统的各种运行参数和需求。

在电力系统频率与有功功率调节中，同步发电机组调速系统是一个基本的工作原理。调速系统通过监测发电机的转速和负载情况，调整机组的有功输出，以维持系统频率的稳定。同时，调速器的特性和工作原理也对频率调节起着重要的作用，需要确保调速器的响应速度和精度，以及与其他系统设备的协调性。

电力系统频率的波动规律与对策是另一个重要内容，需要根据不同时间尺度的频率波动，制定相应的调频调控策略。随机小波动可以通过调速器自动调节来实现一次调频；计划外负荷变化需要调整调速器的功率给定来实现二次调频；而长时间尺度的频率波动需要预测负荷并制定合理的发电计划来实现频率的稳定。

总的来说，电力系统自动化中的频率与有功功率调节涉及到多个方面，包括调速系统的基本原理和特性、频率调节的意义和特点、以及不同时间尺度下的频率波动规律与对策。通过科学合理地进行频率调节，可以提高电力系统的运行效率和可靠性，确保系统稳定运行。因此，电力系统自动化中的频率与有功功率调节是一个至关重要的环节，需要引起相关工程师和研究人员的高度重视和深入研究。管理建模和仿真的文件管理Boualem Benatallah引用此版本：布阿利姆·贝纳塔拉。管理建模和仿真。约瑟夫-傅立叶大学-格勒诺布尔第一大学，1996年。法语。NNT：电话：00345357HAL ID：电话：00345357https://theses.hal.science/tel-003453572008年12月9日提交HAL是一个多学科的开放存取档案馆，用于存放和传播科学研究论文，无论它们是否被公开。论文可以来自法国或国外的教学和研究机构，也可以来自公共或私人研究中心。L’archive ouverte pluridisciplinaire Revit API中的定制化界面开发 # 1. Revit API简介

## 1.1 什么是Revit API

Revit API是Autodesk公司提供的一组应用程序接口，旨在帮助开发人员扩展Revit软件的功能。通过Revit API，开发人员可以创建定制化的界面、实现自动化任务以及与Revit模型进行交互。

## 1.2 Revit API的优势和应用场景

Revit API的优势在于提供了丰富的功能和灵活性，使得开发人员能够根据特定需求定制化界面，提高工作效率。应用场景包括但不限于自动化流程、创建特定定制功能、与其他软件集成等。

## 1.3 Revit API与其他BIM工具的对比

与其他BIM工具相比，R java实现计算机随机出题：随机出十道小学四则运算题目，要求操作数为10以内，运算加、减、乘、除，并可以根据用户的答案给出“√”或者“×”提示，最后计算总得分。好的，以下是 Java 代码实现这个功能：

```java

import java.util.Random;

import java.util.Scanner;

public class RandomArithmeticQuiz {

public static void main(String[] args) {

Scanner scanner = new Scanner(System.in);

Random random = new Random();

int score = 0;

for (int i = 0; i 创意风格PPT模板 (36).pptx MOMODA PowerPoint模板展示了一种创意风格，设计简洁大方，展示内容清晰明了。每一页PPT模板都呈现出一种独特的风格，让演示内容更加生动有趣。Morbi dignissim nisl in diam sagittis, id dapibus nulla pretium. Sed vitae quam metus. Etiam fermentum turpis vel lectus dignissim consectetur. 这些PPT模板适合各种场合使用，无论是商务演示还是教育培训，都能够轻松展示您的内容。

其中的"创意风格PPT模板 (36).pptx"展示了一种现代化的设计风格，色彩鲜明，图标清晰，让您的演示更加吸引人。每一页PPT模板都经过精心设计，呈现出不同的风格和主题，帮助您快速搭建出精美的演示文稿。MOMODA PowerPoint模板也为您提供了丰富的元素和图标，让您在演示中添加更多的细节和亮点，提升演示效果。

与传统的PPT模板不同，MOMODA PowerPoint模板注重创意和个性，让您的演示更加与众不同。Morbi dignissim nisl in diam sagittis, id dapibus nulla pretium. This slide is perfect for long text descriptions Contrary to popular belief, Lorem Ipsum is not simply random text. It has roots in a piece of classical Latin litera; 这些PPT模板不仅适合展示文字和图片，还可轻松添加动画和过渡效果，让您的演示更加生动和引人注目。

总的来说，MOMODA PowerPoint模板是您打造精美演示的最佳选择，无论是商务报道还是学术讲座，都能够帮助您展现出专业的形象和内容。快来下载使用MOMODA PowerPoint模板，让您的演示更上一层楼！

应用于噪声环境下语种识别的GFCC改进算法

首页

期刊介绍

期刊在线

优先发表

当期目录

过刊浏览

浏览排行

下载排行

被引排行

高级检索

编委会

期刊订阅

投稿指南

获奖情况

数据库收录

历史名人

联系我们

所有

标题

作者

关键词

摘要

DOI

栏目

地址

基金

中图分类号

PACS

EEACC

首页

期刊介绍

期刊在线

优先发表

当期目录

过刊浏览

浏览排行

下载排行

被引排行

高级检索

编委会

期刊订阅

投稿指南

获奖情况

数据库收录

历史名人

联系我们

周大春, 邵玉斌, 张昊阁, 龙华, 彭艺. 应用于噪声环境下语种识别的GFCC改进算法[J]. 云南大学学报(自然科学版). doi: 10.7540/j.ynu.20220531

引用本文:

周大春, 邵玉斌, 张昊阁, 龙华, 彭艺. 应用于噪声环境下语种识别的GFCC改进算法[J]. 云南大学学报(自然科学版). doi: 10.7540/j.ynu.20220531

ZHOU Da-chun, SHAO Yu-bin, ZHANG Hao-ge, LONG Hua, PENG Yi. An improved GFCC algorithm for language recognition in noisy environments[J]. Journal of Yunnan University: Natural Sciences Edition. DOI: 10.7540/j.ynu.20220531

Citation:

应用于噪声环境下语种识别的GFCC改进算法

周大春,

邵玉斌,

张昊阁,

龙华,

彭艺

An improved GFCC algorithm for language recognition in noisy environments

ZHOU Da-chun,

SHAO Yu-bin,

ZHANG Hao-ge,

LONG Hua,

PENG Yi

摘要

HTML全文

图(5)

表(2)

参考文献(17)

施引文献

资源附件(0)

摘要

摘要:

不同的噪声在频谱上有不同的特点，使得自动语种识别的性能在噪声环境下显著下降. 针对该问题，提出一种基于改进时域伽马通滤波器倒谱系数(gammatone filter cepstral coefficient, GFCC)特征的语种识别方法. 首先，提取不同噪声背景下的训练集的时域GFCC特征；然后，利用Fisher比计算特征各维对区分语种的相对贡献度大小，分析不同噪声对时域GFCC特征各维的影响，并根据分析来设计合适的权值对特征各维加权，得到语种区分性更强的特征集；最后，利用高斯混合−通用背景模型作为基线系统进行语种识别，以测试所提方法性能. 实验结果表明，在单一噪声背景，信噪比为−5 dB，噪声源分别为粉红噪声、餐厅噪声的条件下，所提方法相比于传统时域GFCC特征方法的识别率分别提升了40.1、20.6个百分点，在其他噪声背景、信噪比下的识别率也有一定程度的提升.

Abstract:

Different noises have different characteristics in the frequency spectrum, which makes the performance of automatic language identification significantly degraded in the noisy environment. To address this problem, a language identification method based on improved time-domain gammatone filter cepstral coefficient (GFCC) features is proposed. First, the time-domain GFCC features are extracted from the training set with different noise backgrounds. Then, the Fisher ratio is used to calculate the relative contribution of each dimension of the features to distinguish languages, to analyse the effect of different noises on each dimension of the time-domain GFCC features, and to design suitable weights to weight each dimension of the features based on the analysis to obtain a feature set with better language discriminatory properties. Finally, a Gaussian mixture model-universal background model is used as the baseline system for language identification to test the performance of the proposed method. The experimental results show that under the conditions of single noise background, signal-to-noise ratio of −5 dB, and noise sources of pink noise and restaurant noise respectively, the identification rate of the proposed method is improved by 40.1 percentage points and 20.6 percentage points respectively compared with the traditional time-domain GFCC feature method, and the identification rate under other noise background and signal-to-noise ratio is also improved to some extent.

HTML全文

参考文献(17)

施引文献

资源附件(0)

下载:

全尺寸图片幻灯片

返回文章

用微信扫码二维码

分享至好友和朋友圈

×Close

导出文件

文件类别

RIS(可直接使用Endnote编辑器进行编辑)

Bib(可直接使用Latex编辑器进行编辑)

Txt

引用内容

引文——仅导出文章的Citation信息

引文和摘要——导出文章的Citation信息和文章摘要信息

导出

关闭

×Close

引用参考文献格式

复制

关闭

Python音频特征提取（MFCC, IMFCC, GFCC, LFCC, PNCC ...） - 知乎

Python音频特征提取（MFCC, IMFCC, GFCC, LFCC, PNCC ...） - 知乎切换模式写文章登录/注册Python音频特征提取（MFCC, IMFCC, GFCC, LFCC, PNCC ...）Drifter?he=human&level=vegetable"--初入音频领域，往往需要学会音频特征提取之后再进一步展开更多的其他工作，然而学会音频处理需要语音信号处理的各种知识（傅里叶变换、DCT变换啊、小波变换啊。。。乱七八糟的我也不懂）但是python有很多第三方库封装好了很多函数使得人们提取更加简单，常见的库（librosa、numpy...）但是这些都还不够，因为我是一点都不懂 hahahha...这里介绍一个库（spafe，直接提取，并且各种特征提取都有jupyter notebook范例）spafe: 简化的Python音频功能提取spafe旨在简化音频中的特征提取。该库涵盖：MFCC，IMFCC，GFCC，LFCC，PNCC，PLP等。它还提供了各种滤波器组模块（Mel，Bark和Gammatone滤波器组）和其他频谱统计信息。Fbank、MFCC、BFCC、GFCC、LFCC、MSRCC、NGCC、PNCC、PSRCC特征提取范例（好多特征都没听过...）spafe库地址安装极其简单： pip install spafe提取各种特征的范例程序地址：这个项目值得更多的Star！！！以下是小补充：范例程序中的fs, sig = scipy.io.wavfile.read("./test.wav")部分音频可能会报错，换成下面这个会好点sig, fs = librosa.load("./test.wav", sr=16000)发布于 2020-04-01 22:33特征提取音频处理音频信息处理赞同 6024 条评论分享喜欢收藏申请

基于MFCC与GFCC混合特征参数的说话人识别

首页

期刊介绍

编委会

投稿指南

期刊订阅

征稿专栏

联系我们

English

应用科学学报 ›› 2019, Vol. 37 ›› Issue (1): 24-32.doi: 10.3969/j.issn.0255-8297.2019.01.003

• 信号与信息处理 •

基于MFCC与GFCC混合特征参数的说话人识别

周萍, 沈昊, 郑凯鹏

桂林电子科技大学电子工程与自动化学院, 广西桂林 541004

收稿日期:2018-02-01

修回日期:2018-04-25

出版日期:2019-01-31

发布日期:2019-01-31

作者简介:周萍,教授,研究方向:语音识别与智能控制,E-mail:940809266@qq.com

基金资助:国家自然科学基金（No.61462017）；广西自然科学基金（No.2014GXNSFAA118353）；广西自动检测技术与仪器重点实验室基金（No.YQ15110）资助

Speaker Recognition Based on Combination of MFCC and GFCC Feature Parameters

ZHOU Ping, SHEN Hao, ZHENG Kai-peng

College of Electric Engineering and Automation, Guilin University of Electronic Technology, Guilin 541004, Guangxi Province, China

Received:2018-02-01

Revised:2018-04-25

Online:2019-01-31

Published:2019-01-31

PDF

261

可视化

摘要/Abstract

摘要：针对说话人识别中单一参数表征不够全面的特点，将抗噪性能一般的传统MFCC参数与鲁棒性更强的GFCC参数相互融合，并结合它们的动态特性构成一种新的混合参数.针对特征参数维数过高造成的冗余，研究了每种特征参数各分量与识别结果的关系，舍弃其中贡献较低的分量以实现特征参数降维的目的，并将混合参数应用于基于高斯混合模型的说话人识别系统.仿真实验表明，该混合特征参数具有更好的识别性能和抗噪性.

关键词:

Mel频率倒谱系数,

混合特征参数,

Gammatone滤波器,

说话人识别

Abstract: Aiming at the issue that single feature parameter of speaker recognition has the shortcoming of low representation ability, a set of mixture feature parameters is formed by combining the single poor anti-noise Mel frequency cepstral coefficients (MFCC) with more robust Gammatone frequency cepstral coefficients (GFCC) and their dynamic differential in this paper. Since the high dimension of the mixture feature parameters, the relationships of each dimension of different feature parameters and recognition results is studied, where dimensionality reduction on high dimensional features is implemented by discarding the dimensions with low contribution ratio. After that, the combination of feature parameters was applied to the speaker recognition system based on Gaussian mixture model. Experimental results show that the combination of parameters can better describe the speakers' feature and have better anti-noise capability.

Key words:

combination of feature parameters,

Mel frequency cepstral coefficients (MFCC),

speaker recognition,

Gammatone filter

中图分类号:

TN912.34

引用本文

周萍, 沈昊, 郑凯鹏. 基于MFCC与GFCC混合特征参数的说话人识别[J]. 应用科学学报, 2019, 37(1): 24-32.

ZHOU Ping, SHEN Hao, ZHENG Kai-peng. Speaker Recognition Based on Combination of MFCC and GFCC Feature Parameters[J]. Journal of Applied Sciences, 2019, 37(1): 24-32.

使用本文

About the GFCC | TheGFCC.org

top of pagePressNewsletterThe GFCCHomeAbout the GFCCWhat the GFCC doesAnnual ReportsGFCC MembersGFCC FellowsThought LeadershipBlogGlobal Competitiveness PrinciplesCall to ActionUniversity and InnovationAdditional ReportsInitiativesDriving Innovation in Times of CrisisUniversity & Research Leadearship ForumGlobal Competitiveness AwardCompetitiveness DecoderFrame the Future SeriesEventsNews and AnnouncementsGFCC CommunityExclusive CommunityGFCC StoriesMoreUse tab to navigate through the menu items.About the GFCCHome/.../About the GFCC/The Global Federation of Competitiveness Councils (GFCC) is a global multi-stakeholder membership organization founded in 2010 with a footprint spanning more than 30 nations. The GFCC is committed to disseminating best practices to accelerate productivity, growth, and prosperity for countries, regions, and cities. The GFCC does that through high-level networking and events, in-depth conversations, advice, and education. For the GFCC, a prosperous society is one driven by innovation, partnership, resilience, inclusiveness, and sustainability.The GFCC’s PurposeAugment the capacity of leaders and economies around the globe to accelerate innovation, build competitiveness, and increase prosperity.The GFCC’s OfferAccess to a highly selective and trustable global network of leaders, world-class knowledge, and accelerated learning.Board of TrusteesChad

HollidayChairman, GFCC | USADeborah Wince-SmithPresident, GFCC | USAOmar

Al-AnsariSecretary General, Qatar Research Development and Innovation Council | QatarRobson Braga de AndradePresident (former), CNI | BrazilIsabel Capeloa GilPresident, UCP | PortugalGeorge HandjinicolaouChairman, Piraeus Bank | GreeceRay O. JohnsonCEO, Technology Innovative Institute | UAEChad EvansExecutive Vice-President, Council on Competitiveness | USARoberto AlvarezExecutive Director, GFCC | USABoad of Trutees 1Board of Trustees 2Board of Trustees 3What members say about the GFCC

PHILLIP PHIRI

EXECUTIVE DIRECTOR | National Competitiveness Commission

The collaborations gained through the GFCC network of experts gave the NCC comprehensive guidance on how government, the private sector, and society need to work together.Click here for the full testimonial.TestimonialThe GFCC community spans35

countriesGFCC members include private sector councils on competitiveness, industry organizations, government agencies, global corporations, and leading research universities. Members bring unique and diverse expertise in various industries, geographies, and knowledge areas. Members commit to paying membership dues yearly to secure their placement in the network and help the organization achieve its mission.

In addition to members, the GFCC network includes over 50 fellows. GFCC fellows are thought leaders and professionals with proven and impactful track records in business, government, and research or civil society organizations. Fellows are part of the GFCC network by invitation only.

The richness and knowledge in the GFCC network are leveraged through global events, initiatives, and member-only opportunities to share best practices.

Learn more about what the GFCC does.Our MembersTo play, press and hold the enter key. To stop, release the enter key.TeamChad

Evans

TreasurerRylie

Pope

Project

CoordinatorDeborah Wince-Smith

PresidentSimone

Melo

Communications

ContractorElaine

Rodriguez

Strategic Management AssociateVanessa

Puerta

Head of

OperationsRoberto

Alvarez

Executive

DirectorBack to home

THE GFCC

Homepage

About the GFCC

What the GFCC does

GFCC Members

GFCC Fellows

THOUGHT LEADERSHIP

Blog

Annual Reports

Annual Principles

Thought Pieces

University & Innovation

INITIATIVES

Driving Innovation in Time of Crises

University & Research Leadearship Forum

Global Competitiveness Award

Frame the Future Series

Competitiveness Decoder

EVENTS

All Events

Past In Person Events

Past Online Events

NEWS & ANNOUNCEMENTS

GFCC COMMUNITY

Global Federation of Competitiveness

Councils

900 17th Street NW, Suite 700

Washington, D.C. 20006

+1 (202) 969 3382

info@thegfcc.orgbottom of page

Global Innovation Summit 2023

Under the theme Shape Tomorrow: Innovate the Sustainable Future, the summit will delve into the creation of business models and policy frameworks that tap into the sustainability opportunity to boost economic prosperity while safeguarding the planet and mitigating the impacts of climate change.

Over four days, experts from business, government agencies, civil society, and research will explore the vital connections between innovation and sustainability, outlining a new pathway to sustainable development by leveraging cutting-edge technologies and forward-looking policy frameworks.

On November 12, the GFCC will host its Annual Meeting, exclusive to its members and fellows. This high-profile gathering is an opportunity to review competitiveness priorities across nations, identify emerging business and social trends, and map relevant investments and projects around the globe. Additionally, members and fellows will reassess the GFCC agenda and discuss future opportunities in the network.

"We must reinvent business models, products, technologies, and policy frameworks to advance sustainability. Innovation is the key for this to happen and requires cross-sector collaboration, which is at the heart of the GFCC and will be a key theme when we meet in Doha", comments GFCC President and founder Deborah L. Wince-Smith.

The summit will also offer unique site visits to innovation hotspots, start-up incubators, higher education institutions, policy organizations, heritage sites, and cultural institutions in Doha.

In 2022, over 400 leaders attended the GIS in Greece, co-hosted by the Delphi Economic Forum and the Council on Competitiveness of Greece. High-profile Greek authorities participated as speakers, including Niki Kerameus, Minister of Education and Religious Affairs, Adonis Georgiadis, Minister of Finance, and Kostas Bakoyannis, Mayor of Athens.

In previous years, the GIS was held in the United States (2010), Brazil (2011), the United Arab Emirates (2012), Korea (2013), Canada (2014), Saudi Arabia (2015), the United Kingdom (2016), Malaysia (2017), Argentina (2018), and Kazakhstan (2019). The 2020 and 2021 editions were online and executed in partnership with the Government of Australia and the Council on Competitiveness.

Visit the GIS website to know more.Explore more GFCC news & announcements

THE GFCC

Homepage

About the GFCC

What the GFCC does

GFCC Members

GFCC Fellows

THOUGHT LEADERSHIP

Blog

Annual Reports

Annual Principles

Thought Pieces

University & Innovation

INITIATIVES

Driving Innovation in Time of Crises

University & Research Leadearship Forum

Global Competitiveness Award

Frame the Future Series

Competitiveness Decoder

EVENTS

All Events

Past In Person Events

Past Online Events

NEWS & ANNOUNCEMENTS

GFCC COMMUNITY

Global Federation of Competitiveness

Councils

900 17th Street NW, Suite 700

Washington, D.C. 20006

+1 (202) 969 3382

info@thegfcc.orgbottom of page

An urgent call to Frame the Future | by The GFCC | Competitive Edge

rgent call to Frame the Future | by The GFCC | Competitive EdgeOpen in appSign upSign inWriteSign upSign inAn urgent call to Frame the FuturePositioning innovation, sustainability, resilience, inclusiveness, and partnerships at the forefrontThe GFCC·FollowPublished inCompetitive Edge·3 min read·Jun 20, 2022--ListenShareBy Roberto AlvarezOne year ago, we started the Frame the Future Conversation Series to discuss how to weave innovation, sustainability, resilience, inclusiveness and partnerships into competitiveness strategies. Since then the world has changed significantly. The Russian invasion of Ukraine shattered the world order, threw Europe in turmoil, and injected complexity into an already challenging global economic landscape.Beyond the humanitarian tragedy, the war has several other effects. It placed additional stress on global supply chains, made energy security a top priority across nations, caused a steep increase in food prices, and contributed to a spike in inflation. It has also thrown Russia back into isolation. Above all, the war cast renewed questions about an already weakened international system.The current state of the world raises questions about our work at the GFCC. What is the role of a global multi-stakeholder organization in this context? More importantly, how relevant is the agenda launched in December 2021? Is the Frame the Future recommendations still relevant in face of the dramatic events and changes the world has experienced lately? Here is my take on these matters.The answer to the first question is clear in my mind. The war has highlighted the importance of the work of the GFCC. In a time of distress, there is a huge need for collaboration platforms that could bring together different countries, sectors, and perspectives. Our organization has a footprint in 33 nations and a unique multi-stakeholder setting. What defines the GFCC is not a sector, an industry, or a political view. But a shared belief in the power of global collaboration to advance national economies and a common desire to be exposed to other realities and work together.In 2021, leaders from 87 countries participated in GFCC activities, and our members and fellows worked together to develop the Frame the Future agenda that is reflected on the report Frame the Future: Guidelines and Recommendations for Future Competitiveness. Is that agenda still relevant? Undoubtfully.Here it is how to connect the current crisis to the priorities we outlined and why it is crucial to advance its implementation.§ Innovation: crises require innovation at speed, and the war calls for innovation in institutions at a global scale to resolve the current and prevent future conflicts.§ Sustainability: the conflict in Ukraine has pushed energy security to center stage, reinforcing the need to invest in local sustainable energy capacity.§ Resilience: the war is threatening food security for many and stressing global supply chains, creating uttermost urgency for economies, societies, and organizations to become more adaptable and resilient in the face of major disruption and disaster.§ Inclusiveness: the conflict shows the importance of digital infrastructure for populations and underlines the need to bring all segments of society into the innovation economy.§ Partnership: crisis response calls for partnerships and tearing down barriers that impede collaboration at all levels between public and private sector.The tragic war in Ukraine is part of an increasingly complex global mosaic and serves as an urgent call to Frame the Future. The agenda that the GFCC launched in December 2021, during the Global Innovation Summit, is more relevant than ever and should be a priority in the years to come.We urge all concerned global stakeholders to check the Frame the Future report and join the GFCC in advancing this relevant agenda. Global leaders have also shared their insights about the future of competitiveness and how to embed sustainability, resilience, inclusiveness, innovation, and partnerships into strategies in an exclusive thought paper series available herehttps://www.thegfcc.org/thought-pieces . I invite you to check these articles.----FollowWritten by The GFCC190 Followers·Editor for Competitive EdgeThe Global Federation of Competitiveness Councils. A network of leaders committed to accelerating global prosperity through fostering innovation ecosystems.FollowHelpStatusAboutCareersBlogPrivacyTermsText to speechTeams

GFCC | The Global Federation of Competitiveness Councils | LinkedIn

Articles

People

Learning

Jobs

Get the app

Join now

GFCC | The Global Federation of Competitiveness Councils

International Affairs

Washington, DC 1,918 followers

Accelerating competitiveness through collaboration

View all 41 employees

Report this company

About us

The Global Federation of Competitiveness Councils is a global network of leaders from competitiveness councils around the world.

The fundamental drivers of national competitiveness are being knitted together in networks that now underpin global economic growth. Innovation, sustainability and resilience—once the foundation for national competitiveness advantage—are now global platforms for prosperity. Acting globally is now a prerequisite to economic competitiveness nationally.

In addition to the goals of sharing best practices among councils and creating a network of global leaders committed to their national prosperity and the prosperity of the world, the Global Federation of Competitiveness Councils will develop and present each year a statement of shared principles for accelerating global competitiveness and prosperity.

Website

http://thegfcc.org/

External link for GFCC | The Global Federation of Competitiveness Councils

Industry

International Affairs

Company size

2-10 employees

Headquarters

Washington, DC

Type

Nonprofit

Locations

Primary

900 17th St NW

Washington, DC 20006, US

Get directions

Employees at GFCC | The Global Federation of Competitiveness Councils

Mark Minevich

Roberto dos Reis Alvarez

Dot Connector | Chief Curiosity Officer

Elaine Rodriguez, Sc.D.

Strategy Fellow at The GFCC (Global Federation of Competitiveness Councils)

Deborah L. Wince-Smith

President & CEO, U.S. Council on Competitiveness

See all employees

Updates

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

11h

Report this post

Do you have a story to tell? Is your organization leading the charge in sustainability, competitiveness, and groundbreaking initiatives?

The GFCC Stories - Innovation Pioneers platform is searching for its next wave of inspiring GFCC members! We want to celebrate YOUR achievements and share them with our extensive network.

Here's what GFCC Stories offers:

↳ Share Your Success: We'll craft a compelling story highlighting your organization's impactful work.

↳ Boost Your Visibility: Reach a global audience through targeted social media promotion and our exclusive network.

↳ Spark Collaboration: Inspire others and foster new connections that lead to exciting partnerships and initiatives.

Your story could be about:

★ A groundbreaking innovation

★ A unique and successful partnership

★ A project making a positive change in your community

Ready to share your story?

Fill out this form: https://lnkd.in/eUa-bbTr, or send an email to info@thegfcc.org

Let's celebrate your achievements and keep pushing the boundaries of innovation together!

Need Inspiration? Check Out These Success Stories:

↳ "Cross-sectoral Initiative in Northern Ireland Produces Biomethane to Achieve Decarbonization," featuring the Center for Competitiveness (CforC) - https://lnkd.in/dwvpmq3d

↳ "Harnessing AI to Bridge the Digital Divide in Eastern Europe," featuring Prof. Razvan Bologa from the Bucharest University of Economic Studies - https://lnkd.in/ehjQ2Vxu

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

In a recent episode of HEDx, Deborah L. Wince-Smith, President & CEO of the US Council on Competitiveness, delivers a compelling message on the critical role of technological innovation in today's economies.

Drawing on historical analogies, Ms. Wince-Smith, who also leads the Universities Research Leadership Forum at the GFCC | The Global Federation of Competitiveness Councils, emphasizes the urgent need for global universities to drive technological transformation.

This is essential to equip future generations with the skills they'll need to thrive in a competitive and productive world.

Tune in to the episode and discover

Why Ms. Wince-Smith believes we've moved from the "Little House on the Prairie" to the "Cyber House on the Prairie."

How universities can spearhead this crucial technological shift.

Practical steps to ensure our economies remain competitive in the face of rapid change.

Don't miss this inspiring conversation! Click below to listen!

EP 107. Lessons from the Bronze Age for university productivity and competitiveness

https://spotify.com

1 Comment

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

We're thrilled to announce that Mr. Hiro Nishiguchi has joined our network as a Distinguished Fellow.

Mr. Nishiguchi brings a wealth of experience working at the intersection of business and policy, having championed innovation and collaboration between corporations and startups in Japan and globally.

A long-time partner of the GFCC, he previously served as a board member during his tenure at the Japan Innovation Network.

Let's give Mr. Nishiguchi a warm welcome in the comments below!

Read all details of this announcement here: https://lnkd.in/eBJPNqpJ

4 Comments

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Edited

Report this post

We have been a strong proponent for the hydrogen economy, recognizing its role in achieving a low-carbon future. Updates

In 2021, we published an interview with Prof. Parker and Dr. Yasushi Sekine, delving into the potential of hydrogen for a greener future and its impact on global competitiveness. You can find it here: https://lnkd.in/gn-HE2g

Now, we will host an Expert Session, exclusive for our Members & Fellows, to discuss the growth of the Hydrogen Economy, on March 14th.

This Expert Session will explore

▶ The rise of hydrogen as a clean and versatile energy carrier for transportation, industry, and power generation.

▶ The global landscape of the hydrogen economy, highlighting opportunities and challenges for different nations.

▶ Showcasing advancements in technology, policy, and investment through ongoing hydrogen projects worldwide.

▶ Insights into financing strategies for hydrogen ventures, including current policy and funding models.

Comment on this post and share your thoughts on the future of the hydrogen economy. Let's start the conversation!

#hydrogeneconomy #renewableenergy #sustainability #greenfuture

Will hydrogen power a net-zero future?

blog.thegfcc.org

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

Our Expert Sessions are opportunities for members to directly connect with on-the-ground experts from various countries. These sessions provide members with valuable information, insights, and even business chances. Help us with the question below.

This content isn’t available here

Get full access in the Linkedin app

Download the app

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

The Aston University is proud to announce the launch of the Aston Digital Futures Institute (ADFI), a new institute dedicated to developing innovative solutions to some of the world's most pressing challenges!

The ADFI brings together experts from a wide range of disciplines, including engineering, business, and computer science, to work collaboratively on projects that have a real-world impact.

The initiative's focus goes from using AI to improve healthcare to tackling climate change and social inequality, and it's led by Professor Aleks Subic, Vice-Chancellor and Chief Executive of Aston, and co-chair of the GFCC University and Research Leadership Forum.

You can read the full GFCC Story here: https://lnkd.in/eniX-AA2

3 Comments

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

Crises are inevitable, but are they also catalysts for positive change?

The DITC, a global initiative exploring the intersection of crisis and innovation, is on a mission to answer that question.

Here's what we've been up to:

Documented key discussions from 2022, exploring historical moments where innovation thrived during adversity.

Welcomed two talented researchers: John E. Katsos and Ailun Gu, who delved deep into the world of crisis definitions and frameworks.

Published a white paper summarizing findings, along with three insightful case studies:

- The 9/11 Attacks

- The COVID-19 Pandemic

- The Armed Conflict in Ukraine

Ready to dive deeper? Head over to our page and access all our reports and case studies: https://lnkd.in/eAyN8hCF

#Crisis #Innovation #CrisisManagement

1 Comment

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

Report this post

Despite facing devastating losses, Ukraine has unveiled remarkable innovation in the face of conflict.

This case study spotlights key examples and explores how the nation can leverage this momentum for a brighter future.

Get the full story and actionable insights: https://ow.ly/uj2Y50QCBme

#Ukraine #Innovation #Resilience #Crisis

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

1mo

Report this post

Early recognition is vital in pandemic response. But are we truly prepared?

Dive into our report to discover innovative solutions and bridge the gap in crisis preparedness.

Download the report: https://ow.ly/zqvQ50QBBig

#Crisis #Innovation

Comment

GFCC | The Global Federation of Competitiveness Councils

1,918 followers

1mo

Report this post

We are delighted to announce Dr. Hippolyte Fofack, Research Associate at Harvard University Center for African Studies, as our new Distinguished Fellow.

With over 20 years of experience in economic policy, development economics, and international finance, Dr. Fofack brings invaluable expertise to our organization.

Formerly at the African Export-Import Bank, Dr. Fofack's contributions to initiatives like the African Continental Free Trade Agreement (AfCFTA) highlight his commitment to advancing economic development in Africa and beyond.

Join us in welcoming Dr. Fofack, and read more about his career here: https://lnkd.in/eEk64x79

11 Comments

Comment

Join now to see what you are missing

Find people you know at GFCC | The Global Federation of Competitiveness Councils

Browse recommended jobs for you

View all updates, news, and articles

Join now

Affiliated pages

The Innovation Nexus

Public Policy Offices

Similar pages

Council on Competitiveness

Public Policy Offices

Washington, DC

Horizon Group

Professional Services

Geneva, Geneve

African Energy Chamber

Oil and Gas

Johannesburg, Gauteng

IFCU-International Federation of Catholic Universities

Education Administration Programs

Paris, Ile de France

TIGER 21

Professional Training and Coaching

New York, NY

GreenMet

Financial Services

Washington, District of Columbia

Marine Forces Reserve

Armed Forces

New Orleans, Louisiana

Centre for Competitiveness

Non-profit Organizations

Belfast, Northern Ireland

CIP - Confederação Empresarial de Portugal

Civic and Social Organizations

Lisboa, Lisboa

West Midlands Combined Authority

Government Administration

Birmingham, West Midlands

Show more similar pages

Show fewer similar pages

Browse jobs

Engineer jobs

608,159 open jobs

Lead jobs

2,280,087 open jobs

Investment Strategist jobs

5,887 open jobs

Executive jobs

700,389 open jobs

Vice President jobs

243,153 open jobs

Editor jobs

23,258 open jobs

Finance Officer jobs

62,417 open jobs

Unity Developer jobs

1,830 open jobs

Director jobs

1,374,979 open jobs

Associate Software Engineer jobs

428,801 open jobs

Administrator jobs

476,673 open jobs

Scrum Master jobs

146,132 open jobs

Developer jobs

344,797 open jobs

Quality Assurance Automation Engineer jobs

8,620 open jobs

Director of Engineering jobs

22,553 open jobs

Chief Executive Officer jobs

186,821 open jobs

Engineering Manager jobs

197,372 open jobs

Product Designer jobs

73,307 open jobs

Designer jobs

81,042 open jobs

PHP Developer jobs

15,381 open jobs

Show more jobs like this

Show fewer jobs like this

Follow

tokenpocket官网苹果下载|gfcc

tokenpocket官网苹果下载|gfcc

Global Federation of Competitiveness Councils

声学感知刻度(mel scale、Bark scale、ERB)与声学特征提取(MFCC、BFCC、GFCC)_hz2erb-CSDN博客

GFCC特征参数改进建议 - CSDN文库

应用于噪声环境下语种识别的GFCC改进算法

Python音频特征提取（MFCC, IMFCC, GFCC, LFCC, PNCC ...） - 知乎

基于MFCC与GFCC混合特征参数的说话人识别

About the GFCC | TheGFCC.org

Global Innovation Summit 2023

An urgent call to Frame the Future | by The GFCC | Competitive Edge

GFCC | The Global Federation of Competitiveness Councils | LinkedIn

最近的新闻

您可能喜欢的文章

苹果怎么下载比特派钱包|idex是什么

tokenpocket钱包官网注册|500万泰铢等于多少人民币

区块链冷钱包开发——安全可靠的加密货币存储

网页下载以太坊钱包，简单易用的数字货币钱包

TP钱包空投币能点吗？

TP钱包国内版和国际版有什么区别？

如何防范TP钱包被盗，保障你的数字资产安全