Python implements the function of a music retriever

Listen to songs and identify songs. As the name suggests, you use the device to "listen" to songs, and then it will tell you what song it is. And nine times out of ten, it will have to play the song for you. Such a function has long appeared in applications such as QQ Music. Today we are going to make our own song recognition by listening to songs

The overall flow chart we designed is very simple:

---- -
Recording part
-----

If we want to "listen", we must first have the recording process. In our experiments, our music library also uses our recording code to record, and then extracts features and stores them in the database. We use the following idea to record

# coding=utf8
import wave

import pyaudio


class recode():
 def recode(self, CHUNK=44100, FORMAT=pyaudio.paInt16, CHANNELS=2, RATE=44100, RECORD_SECONDS=200,
    WAVE_OUTPUT_FILENAME="record.wav"):
  '''

  :param CHUNK: 缓冲区大小
  :param FORMAT: 采样大小
  :param CHANNELS:通道数
  :param RATE:采样率
  :param RECORD_SECONDS:录的时间
  :param WAVE_OUTPUT_FILENAME:输出文件路径
  :return:
  '''
  p = pyaudio.PyAudio()
  stream = p.open(format=FORMAT,
      channels=CHANNELS,
      rate=RATE,
      input=True,
      frames_per_buffer=CHUNK)
  frames = []
  for i in range(0, int(RATE / CHUNK * RECORD_SECONDS)):
   data = stream.read(CHUNK)
   frames.append(data)
  stream.stop_stream()
  stream.close()
  p.terminate()
  wf = wave.open(WAVE_OUTPUT_FILENAME, 'wb')
  wf.setnchannels(CHANNELS)
  wf.setsampwidth(p.get_sample_size(FORMAT))
  wf.setframerate(RATE)
  wf.writeframes(''.join(frames))
  wf.close()


if __name__ == '__main__':
 a = recode()
 a.recode(RECORD_SECONDS=30, WAVE_OUTPUT_FILENAME='record_pianai.wav')

What is the form of the song we have recorded?

If you only look at one channel, it is a one-dimensional array, which probably looks like this

We assign it according to the index value Drawn as the horizontal axis, it is the form of audio we often see.

Audio processing part

We are going to write our core code here. The crucial "how to identify a song". Think about how we humans differentiate between songs? Is it by thinking of a one-dimensional array like the one above? Is it based on the loudness of the song? neither.

We memorize songs through the sequence of unique frequencies heard by the ears, so if we want to write about listening to songs and identifying songs, we have to make a fuss about the frequency sequence of the audio.

Review what is Fourier transform. The blogger's "Signals and Systems" class was very popular, but although I didn't write down the specific transformation form in the class, I still had a perceptual understanding.

The essence of Fourier transform is to transform the time domain signal into the frequency domain signal. That is to say, the original X and Y axes were our array subscripts and array elements respectively, but now they become the frequency (this is not accurate, but it is understood here correctly) and the component size at this frequency.

How to understand the frequency domain? For those of us who don’t know much about signal processing, the most important thing is to change our understanding of the composition of audio. We originally thought that audio is like the waveform we gave at the beginning, which has an amplitude at each time, and different amplitude sequences constitute our specific sound. Now, we think that sound is a mixture of different frequency signals, and each of their signals exists from beginning to end. And they contribute according to their projected components.

Let’s see what it looks like to convert a song into the frequency domain?

We can observe that the components of these frequencies are not average, and the differences are very large. To a certain extent, we can think that the obviously raised peak in the picture is a frequency signal with large output energy, which means that this signal occupies a high position in this audio frequency. So we chose such a signal to extract the characteristics of the song.

But don’t forget, what we talked about before was a frequency sequence. With a set of Fourier transforms, we can only know the frequency information of the entire song, and then we lose the time relationship. We say There is no way to talk about the "sequence". So we adopted a more compromise method, dividing the audio into small chunks according to time. Here I divided it into 40 chunks per second.

Leave a question here: Why use small blocks instead of one large block per second?

We perform Fourier transform on each block, and then modulo it to obtain arrays. We take the subscript with the largest module length in the four intervals with subscript values ​​​​(0,40), (40,80), (80,120), (120,180), and synthesize a four-tuple. This is our core Audio "fingerprint".

The "fingerprint" we extracted is similar to the following

(39, 65, 110, 131), (15, 66, 108, 161), (3, 63, 118, 146), (11, 62, 82, 158), (15, 41, 95, 140), (2, 71, 106, 143), (15, 44, 80, 133), (36, 43, 80, 135), (22, 58, 80, 120), (29, 52, 89, 126), (15, 59, 89, 126), (37, 59, 89, 126), (37, 59, 89, 126), (37, 67, 119, 126)

音频处理的类有三个方法：载入数据，傅里叶变换，播放音乐。
如下：

# coding=utf8
import os
import re
import wave
import numpy as np
import pyaudio


class voice():
 def loaddata(self, filepath):
  '''

  :param filepath: 文件路径，为wav文件
  :return: 如果无异常则返回True，如果有异常退出并返回False
  self.wave_data内储存着多通道的音频数据，其中self.wave_data[0]代表第一通道
  具体有几通道，看self.nchannels
  '''
  if type(filepath) != str:
   print 'the type of filepath must be string'
   return False
  p1 = re.compile('\.wav')
  if p1.findall(filepath) is None:
   print 'the suffix of file must be .wav'
   return False
  try:
   f = wave.open(filepath, 'rb')
   params = f.getparams()
   self.nchannels, self.sampwidth, self.framerate, self.nframes = params[:4]
   str_data = f.readframes(self.nframes)
   self.wave_data = np.fromstring(str_data, dtype=np.short)
   self.wave_data.shape = -1, self.sampwidth
   self.wave_data = self.wave_data.T
   f.close()
   self.name = os.path.basename(filepath) # 记录下文件名
   return True
  except:
   print 'File Error!'

 def fft(self, frames=40):
  '''
  :param frames: frames是指定每秒钟分块数
  :return:
  '''
  block = []
  fft_blocks = []
  self.high_point = []
  blocks_size = self.framerate / frames # block_size为每一块的frame数量
  blocks_num = self.nframes / blocks_size # 将音频分块的数量
  for i in xrange(0, len(self.wave_data[0]) - blocks_size, blocks_size):
   block.append(self.wave_data[0][i:i + blocks_size])
   fft_blocks.append(np.abs(np.fft.fft(self.wave_data[0][i:i + blocks_size])))
   self.high_point.append((np.argmax(fft_blocks[-1][:40]),
         np.argmax(fft_blocks[-1][40:80]) + 40,
         np.argmax(fft_blocks[-1][80:120]) + 80,
         np.argmax(fft_blocks[-1][120:180]) + 120,
         # np.argmax(fft_blocks[-1][180:300]) + 180,
         )) # 提取指纹的关键步骤，没有取最后一个，但是保留了这一项，可以想想为什么去掉了？

 def play(self, filepath):
  '''
  用来做音频播放的方法
  :param filepath:文件路径 
  :return: 
  '''
  chunk = 1024
  wf = wave.open(filepath, 'rb')
  p = pyaudio.PyAudio()
  # 打开声音输出流
  stream = p.open(format=p.get_format_from_width(wf.getsampwidth()),
      channels=wf.getnchannels(),
      rate=wf.getframerate(),
      output=True)
  # 写声音输出流进行播放
  while True:
   data = wf.readframes(chunk)
   if data == "":
    break
   stream.write(data)

  stream.close()
  p.terminate()


if __name__ == '__main__':
 p = voice()
 p.loaddata('record_beiyiwang.wav')
 p.fft()

这里面的self.high_point是未来应用的核心数据。列表类型，里面的元素都是上面所解释过的指纹的形式。

数据存储和检索部分

因为我们是事先做好了曲库来等待检索，所以必须要有相应的持久化方法。我采用的是直接用mysql数据库来存储我们的歌曲对应的指纹，这样有一个好处：省写代码的时间

我们将指纹和歌曲存成这样的形式：
顺便一说：为什么各个歌曲前几个的指纹都一样？（当然，后面肯定是千差万别的）其实是音乐开始之前的时间段中没有什么能量较强的点，而由于我们44100的采样率比较高，就会导致开头会有很多重复，别担心。

我们怎么来进行匹配呢？我们可以直接搜索音频指纹相同的数量，不过这样又损失了我们之前说的序列，我们必须要把时间序列用上。否则一首歌曲越长就越容易被匹配到，这种歌曲像野草一样疯狂的占据了所有搜索音频的结果排行榜中的第一名。而且从理论上说，音频所包含的信息就是在序列中体现，就像一句话是靠各个短语和词汇按照一定顺序才能表达出它自己的意思。单纯的看两个句子里的词汇重叠数是完全不能判定两句话是否相似的。我们采用的是下面的算法，不过我们这只是实验性的代码，算法设计的很简单，效率不高。建议想要做更好的结果的同学可以使用改进的DTW算法。

我们在匹配过程中滑动指纹序列，每次比对模式串和源串的对应子串，如果对应位置的指纹相同，则这次的比对相似值加一，我们把滑动过程中得到的最大相似值作为这两首歌的相似度。

举例：

曲库中的一首曲子的指纹序列:[fp13, fp20, fp10, fp29, fp14, fp25, fp13, fp13, fp20, fp33, fp14]

检索音乐的指纹序列: [fp14, fp25, fp13, fp17]

比对过程：


最终的匹配相似值为3

存储检索部分的实现代码

# coding=utf-8

import os

import MySQLdb

import my_audio


class memory():
 def __init__(self, host, port, user, passwd, db):
  '''
  初始化存储类
  :param host:主机位置
  :param port:端口
  :param user:用户名
  :param passwd:密码
  :param db:数据库名
  '''
  self.host = host
  self.port = port
  self.user = user
  self.passwd = passwd
  self.db = db

 def addsong(self, path):
  '''
  添加歌曲方法，将指定路径的歌曲提取指纹后放到数据库
  :param path:路径
  :return:
  '''
  if type(path) != str:
   print 'path need string'
   return None
  basename = os.path.basename(path)
  try:
   conn = MySQLdb.connect(host=self.host, port=self.port, user=self.user, passwd=self.passwd, db=self.db,
         charset='utf8')
   # 创建与数据库的连接
  except:
   print 'DataBase error'
   return None
  cur = conn.cursor()
  namecount = cur.execute("select * from fingerprint.musicdata WHERE song_name = '%s'" % basename)
  # 查询新添加的歌曲是否已经在曲库中了
  if namecount > 0:
   print 'the song has been record!'
   return None
  v = my_audio.voice()
  v.loaddata(path)
  v.fft()
  cur.execute("insert into fingerprint.musicdata VALUES('%s','%s')" % (basename, v.high_point.__str__()))
  # 将新歌曲的名字和指纹存到数据库中
  conn.commit()
  cur.close()
  conn.close()

 def fp_compare(self, search_fp, match_fp):
  '''
  指纹比对方法。
  :param search_fp: 查询指纹
  :param match_fp: 库中指纹
  :return:最大相似值 float
  '''
  if len(search_fp) > len(match_fp):
   return 0
  max_similar = 0
  search_fp_len = len(search_fp)
  match_fp_len = len(match_fp)
  for i in range(match_fp_len - search_fp_len):
   temp = 0
   for j in range(search_fp_len):
    if match_fp[i + j] == search_fp[j]:
     temp += 1
   if temp > max_similar:
    max_similar = temp
  return max_similar

 def search(self, path):
  '''
  从数据库检索出
  :param path: 需要检索的音频的路径
  :return:返回列表，元素是二元组，第一项是匹配的相似值，第二项是歌曲名
  '''
  v = my_audio.voice()
  v.loaddata(path)
  v.fft()
  try:
   conn = MySQLdb.connect(host=self.host, port=self.port, user=self.user, passwd=self.passwd, db=self.db,
         charset='utf8')
  except:
   print 'DataBase error'
   return None
  cur = conn.cursor()
  cur.execute("SELECT * FROM fingerprint.musicdata")
  result = cur.fetchall()
  compare_res = []
  for i in result:
   compare_res.append((self.fp_compare(v.high_point[:-1], eval(i[1])), i[0]))
  compare_res.sort(reverse=True)
  cur.close()
  conn.close()
  print compare_res
  return compare_res

 def search_and_play(self, path):
  '''
  跟上个方法一样，不过增加了将搜索出的最优结果直接播放的功能
  :param path: 带检索歌曲路径
  :return:
  '''
  v = my_audio.voice()
  v.loaddata(path)
  v.fft()
  # print v.high_point
  try:
   conn = MySQLdb.connect(host=self.host, port=self.port, user=self.user, passwd=self.passwd, db=self.db,
         charset='utf8')
  except:
   print 'DataBase error'
   return None
  cur = conn.cursor()
  cur.execute("SELECT * FROM fingerprint.musicdata")
  result = cur.fetchall()
  compare_res = []
  for i in result:
   compare_res.append((self.fp_compare(v.high_point[:-1], eval(i[1])), i[0]))
  compare_res.sort(reverse=True)
  cur.close()
  conn.close()
  print compare_res
  v.play(compare_res[0][1])
  return compare_res


if __name__ == '__main__':
 sss = memory('localhost', 3306, 'root', 'root', 'fingerprint')
 sss.addsong('taiyangzhaochangshengqi.wav')
 sss.addsong('beiyiwangdeshiguang.wav')
 sss.addsong('xiaozezhenger.wav')
 sss.addsong('nverqing.wav')
 sss.addsong('the_mess.wav')
 sss.addsong('windmill.wav')
 sss.addsong('end_of_world.wav')
 sss.addsong('pianai.wav')

 sss.search_and_play('record_beiyiwang.wav')

总结

我们这个实验很多地方都很粗糙，核心的算法是从shazam公司提出的算法吸取的“指纹”的思想。希望读者可以提出宝贵建议。

更多Python implements the function of a music retriever相关文章请关注PHP中文网！