语音中的MFCC和Fbank

的区别是什么？

MFCC（Mel-frequency cepstral coefficients）和Fbank（filter banks）都是音频信号处理中常用的特征提取方法。其中，MFCC在人类听觉模型中模拟了耳蜗的工作原理，将频率转换成Mel频率，再采用离散余弦变换（DCT）将信号投影到一个低维的空间中，得到一组提取出来的特征向量。而Fbank则是将音频信号通过一组不同宽度的三角滤波器进行滤波，再通过一些统计特征比如平均值、标准差等得到相应的特征向量。因此，两者的主要区别在于信号的转换方式和特征提取方式不同。

相关问题

python实现mfcc

MFCC (Mel Frequency Cepstral Coefficients，梅尔频率倒谱系数) 是一种常用的音频特征提取方法，可以用于语音识别、音乐信息检索等领域。下面是一个简单的 Python 实现：

首先，我们需要导入一些库：

import numpy as np
from scipy.fftpack import dct
from scipy import signal

接下来，定义一个函数来实现 MFCC 特征提取：

def mfcc(signal, samplerate=16000, winlen=0.025, winstep=0.01, numcep=13, nfilt=26, nfft=512, lowfreq=0, highfreq=None, preemph=0.97, ceplifter=22):
    """
    Compute MFCC features from an audio signal.
    """
    # Pre-emphasis
    signal = signal - np.mean(signal)
    signal = np.append(signal[0], signal[1:] - preemph * signal[:-1])
    
    # Framing
    winlen = int(winlen * samplerate)
    winstep = int(winstep * samplerate)
    frames = signal[:len(signal) - winlen: winstep] * np.hamming(winlen)
    
    # Power spectrum
    pspec = np.abs(np.fft.rfft(frames, nfft))**2
    if highfreq is None:
        highfreq = samplerate / 2
    else:
        highfreq = min(highfreq, samplerate / 2)
    binfreq = np.fft.rfftfreq(nfft, 1/samplerate)
    bins = np.arange(nfilt + 2) / (nfilt + 1) * (highfreq - lowfreq) + lowfreq
    bwidth = bins[1:] - bins[:-1]
    
    # Filterbank
    fbank = np.zeros((nfilt, int(np.floor(nfft / 2 + 1))))
    for i in range(nfilt):
        low = bins[i]
        center = bins[i + 1]
        high = bins[i + 2]
        lbin = np.floor(low * nfft / samplerate) + 1
        cbin = np.floor(center * nfft / samplerate) + 1
        hbin = np.floor(high * nfft / samplerate) + 1
        fbank[i, int(lbin): int(cbin)] = (np.arange(cbin, int(lbin) - 1, -1) - lbin) / (cbin - lbin)
        fbank[i, int(cbin): int(hbin)] = (hbin - np.arange(cbin, hbin)) / (hbin - cbin)
    
    # Apply filterbank
    feat = np.dot(pspec, fbank.T)
    feat = np.where(feat == 0, np.finfo(float).eps, feat)
    feat = np.log(feat)
    
    # DCT
    feat = dct(feat, type=2, axis=1, norm='ortho')[:, :numcep]
    
    # Cepstral lifter
    lifter = 1 + (ceplifter / 2) * np.sin(np.pi * np.arange(numcep) / ceplifter)
    feat = feat * lifter
    
    return feat

其中，`signal` 是输入的音频信号，`samplerate` 是采样率，`winlen` 是窗口长度，`winstep` 是窗口步长，`numcep` 是 MFCC 的维度，`nfilt` 是滤波器组数，`nfft` 是 FFT 的长度，`lowfreq` 和 `highfreq` 是滤波器组的频率范围，`preemph` 是预加重系数，`ceplifter` 是 cepstral lifter 系数。

该函数的返回值是一个二维数组，每行表示一个音频帧的 MFCC 特征。你可以将这些特征作为输入用于下游任务，比如说语音识别。

python实现MFCC计算

MFCC(Mel Frequency Cepstral Coefficients)是一种常用于语音识别和音频处理的特征提取方法，可以将音频信号转化为一组能够描述其特征的系数。以下是Python实现MFCC计算的示例代码：

import librosa
import numpy as np

# Load audio file
y, sr = librosa.load('audio.wav', sr=16000)

# Pre-emphasis
pre_emphasis = 0.97
emphasized_signal = np.append(y[0], y[1:] - pre_emphasis * y[:-1])

# Framing
frame_size = 0.025  # 25ms
frame_stride = 0.01  # 10ms
frame_length, frame_step = frame_size * sr, frame_stride * sr
signal_length = len(emphasized_signal)
frame_length = int(round(frame_length))
frame_step = int(round(frame_step))
num_frames = int(np.ceil(float(np.abs(signal_length - frame_length)) / frame_step))

pad_signal_length = num_frames * frame_step + frame_length
z = np.zeros((pad_signal_length - signal_length))
pad_signal = np.append(emphasized_signal, z)

indices = np.tile(np.arange(0, frame_length), (num_frames, 1)) + np.tile(
    np.arange(0, num_frames * frame_step, frame_step), (frame_length, 1)).T
frames = pad_signal[indices.astype(np.int32, copy=False)]

# Windowing
frames *= np.hamming(frame_length)

# Fourier Transform and Power Spectrum
NFFT = 512
mag_frames = np.absolute(np.fft.rfft(frames, NFFT))  # Magnitude of the FFT
pow_frames = ((1.0 / NFFT) * ((mag_frames) ** 2))  # Power Spectrum

# Filter Banks
nfilt = 40
low_freq_mel = 0
high_freq_mel = (2595 * np.log10(1 + (sr / 2) / 700))  # Convert Hz to Mel
mel_points = np.linspace(low_freq_mel, high_freq_mel, nfilt + 2)  # Equally spaced in Mel scale
hz_points = (700 * (10 ** (mel_points / 2595) - 1))  # Convert Mel to Hz
bin = np.floor((NFFT + 1) * hz_points / sr)

fbank = np.zeros((nfilt, int(np.floor(NFFT / 2 + 1))))
for m in range(1, nfilt + 1):
    f_m_minus = int(bin[m - 1])  # left
    f_m = int(bin[m])  # center
    f_m_plus = int(bin[m + 1])  # right

    for k in range(f_m_minus, f_m):
        fbank[m - 1, k] = (k - bin[m - 1]) / (bin[m] - bin[m - 1])
    for k in range(f_m, f_m_plus):
        fbank[m - 1, k] = (bin[m + 1] - k) / (bin[m + 1] - bin[m])

# Apply filter banks
filter_banks = np.dot(pow_frames, fbank.T)
filter_banks = np.where(filter_banks == 0, np.finfo(float).eps, filter_banks)  # Numerical Stability
filter_banks = 20 * np.log10(filter_banks)  # dB

# MFCCs
num_ceps = 12
mfcc = dct(filter_banks, type=2, axis=1, norm='ortho')[:, 1 : (num_ceps + 1)]

# Mean normalization
mfcc -= (np.mean(mfcc, axis=0) + 1e-8)

print(mfcc)

这里使用了Librosa库来加载音频文件和进行预加重，以及进行帧划分、加窗、傅里叶变换、滤波器组应用和离散余弦变换等步骤。最后，对MFCC系数进行均值归一化。