如何在Python中获取BPM和节奏音频特征

Question

我正在参与一个项目，需要提取一些歌曲的特征，比如每分钟的节拍数（BPM）、节奏等等。不过，我还没有找到一个合适的Python库，能够准确地检测这些特征。

有没有人能给点建议？

（在Matlab中，我知道有一个叫Mirtoolbox的项目，可以在处理本地的mp3文件后提供BPM和节奏的信息。）

Answer 1

我在这里找到了一段代码，作者是@scaperot，可能对你有帮助：点击这里查看

import wave, array, math, time, argparse, sys
import numpy, pywt
from scipy import signal
import pdb
import matplotlib.pyplot as plt

def read_wav(filename):

    #open file, get metadata for audio
    try:
        wf = wave.open(filename,'rb')
    except IOError, e:
        print e
        return

    # typ = choose_type( wf.getsampwidth() ) #TODO: implement choose_type
    nsamps = wf.getnframes();
    assert(nsamps > 0);

    fs = wf.getframerate()
    assert(fs > 0)

    # read entire file and make into an array
    samps = list(array.array('i',wf.readframes(nsamps)))
    #print 'Read', nsamps,'samples from', filename
    try:
        assert(nsamps == len(samps))
    except AssertionError, e:
        print  nsamps, "not equal to", len(samps)

    return samps, fs

# print an error when no data can be found
def no_audio_data():
    print "No audio data for sample, skipping..."
    return None, None

# simple peak detection
def peak_detect(data):
    max_val = numpy.amax(abs(data)) 
    peak_ndx = numpy.where(data==max_val)
    if len(peak_ndx[0]) == 0: #if nothing found then the max must be negative
        peak_ndx = numpy.where(data==-max_val)
    return peak_ndx

def bpm_detector(data,fs):
    cA = [] 
    cD = []
    correl = []
    cD_sum = []
    levels = 4
    max_decimation = 2**(levels-1);
    min_ndx = 60./ 220 * (fs/max_decimation)
    max_ndx = 60./ 40 * (fs/max_decimation)

    for loop in range(0,levels):
        cD = []
        # 1) DWT
        if loop == 0:
            [cA,cD] = pywt.dwt(data,'db4');
            cD_minlen = len(cD)/max_decimation+1;
            cD_sum = numpy.zeros(cD_minlen);
        else:
            [cA,cD] = pywt.dwt(cA,'db4');
        # 2) Filter
        cD = signal.lfilter([0.01],[1 -0.99],cD);

        # 4) Subtractargs.filename out the mean.

        # 5) Decimate for reconstruction later.
        cD = abs(cD[::(2**(levels-loop-1))]);
        cD = cD - numpy.mean(cD);
        # 6) Recombine the signal before ACF
        #    essentially, each level I concatenate 
        #    the detail coefs (i.e. the HPF values)
        #    to the beginning of the array
        cD_sum = cD[0:cD_minlen] + cD_sum;

    if [b for b in cA if b != 0.0] == []:
        return no_audio_data()
    # adding in the approximate data as well...    
    cA = signal.lfilter([0.01],[1 -0.99],cA);
    cA = abs(cA);
    cA = cA - numpy.mean(cA);
    cD_sum = cA[0:cD_minlen] + cD_sum;

    # ACF
    correl = numpy.correlate(cD_sum,cD_sum,'full') 

    midpoint = len(correl) / 2
    correl_midpoint_tmp = correl[midpoint:]
    peak_ndx = peak_detect(correl_midpoint_tmp[min_ndx:max_ndx]);
    if len(peak_ndx) > 1:
        return no_audio_data()

    peak_ndx_adjusted = peak_ndx[0]+min_ndx;
    bpm = 60./ peak_ndx_adjusted * (fs/max_decimation)
    print bpm
    return bpm,correl


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Process .wav file to determine the Beats Per Minute.')
    parser.add_argument('--filename', required=True,
                   help='.wav file for processing')
    parser.add_argument('--window', type=float, default=3,
                   help='size of the the window (seconds) that will be scanned to determine the bpm.  Typically less than 10 seconds. [3]')

    args = parser.parse_args()
    samps,fs = read_wav(args.filename)

    data = []
    correl=[]
    bpm = 0
    n=0;
    nsamps = len(samps)
    window_samps = int(args.window*fs)         
    samps_ndx = 0;  #first sample in window_ndx 
    max_window_ndx = nsamps / window_samps;
    bpms = numpy.zeros(max_window_ndx)

    #iterate through all windows
    for window_ndx in xrange(0,max_window_ndx):

        #get a new set of samples
        #print n,":",len(bpms),":",max_window_ndx,":",fs,":",nsamps,":",samps_ndx
        data = samps[samps_ndx:samps_ndx+window_samps]
        if not ((len(data) % window_samps) == 0):
            raise AssertionError( str(len(data) ) ) 

        bpm, correl_temp = bpm_detector(data,fs)
        if bpm == None:
            continue
        bpms[window_ndx] = bpm
        correl = correl_temp

        #iterate at the end of the loop
        samps_ndx = samps_ndx+window_samps;
        n=n+1; #counter for debug...

    bpm = numpy.median(bpms)
    print 'Completed.  Estimated Beats Per Minute:', bpm

    n = range(0,len(correl))
    plt.plot(n,abs(correl)); 
    plt.show(False); #plot non-blocking
    time.sleep(10);
plt.close();

Answer 2

这个回答虽然晚了一年，但还是记录一下。我找到了三个可以用Python调用的音频库，它们可以从音频中提取特征。安装这些库并不简单，因为它们实际上是用C语言写的，你需要正确编译Python的接口，并把它们添加到路径中才能导入。不过，这里是它们的链接：

• Yaafe
• Aubio
• LibXtract
• Essentia

Answer 3

你要找的是Echo Nest API：

http://echonest.github.io/remix/

Python的接口功能很强大，不过安装Echo Nest可能会有点麻烦，因为他们的团队似乎没有做出很好的安装程序。

不过，它并不支持本地处理。相反，它会计算音频指纹，然后把歌曲上传到Echo Nest的服务器，利用他们不公开的算法来提取信息。