使用机器学习方法识别音频文件的音乐和演讲

背景

最近下载了一批类似百家讲坛的音频文件。这些文件前面部分是演讲类的音频，主要讲历史的，后面一部分是音乐。
但是我只想听演讲类部分，不想听音乐。所以希望把文件切割，把音乐部分切走，只留下演讲部分。
观察文件，发现每个文件的音乐都不一样，演讲和音乐的长度也不一样。
这里一个技术难点就是怎么识别哪些音频是演讲，哪些音频是音乐。
通过KNN算法，1s的音频文件的预测正确率是92%。
同时3s都判断为音乐才进行分割，整个文件的分割正确率是98%。

音频文件和源码

音频文件和源码可以在这里下载

一、把音频文件转换为数字

# encoding=gbk
import random
import wave
import matplotlib.pyplot as plt
import numpy as np
import os

# nchannels 声道
# sampwidth 样本宽度
# framerate 帧率，也就是一秒有多少帧
# nframes 文件一共有多少帧

def pre_deal(file_path):
    """音频解析，返回音频数据"""
    f = wave.open(file_path, 'rb')
    params = f.getparams()
    nchannels, sampwidth, framerate, nframes = params[:4]
    strData = f.readframes(nframes)  # 读取音频，字符串格式
    waveData = np.fromstring(strData, dtype=np.int16)  # 将字符串转化为int

    waveData = waveData[::nchannels]  # 根据声道数，转换为单声道
    rate = 20.00
    framerate = framerate / rate  # 降低帧率
    nframes = nframes / rate  # 降低帧率
    waveData = waveData[::int(rate)]

    # wave幅值归一化
    max_ = float(max(abs(waveData)))
    waveData = waveData / max_

    return waveData, framerate, nframes


def plpot(waveData):
    """画图"""
    time = [i for i, v in enumerate(waveData)]
    plt.plot(time, waveData)
    plt.xlabel("Time")
    plt.ylabel("Amplitude")
    plt.title("Single channel wavedata")
    plt.grid('on')  # 标尺，on：有，off:无。
    plt.show()


def mp3towav(file_path, to_file_path):
    """mp3文件转wav文件"""
    if os.path.exists(to_file_path):
        return to_file_path
    from pydub import AudioSegment
    print file_path
    song1 = AudioSegment.from_mp3(file_path)
    song1.export(to_file_path, 'wav')
    return to_file_path


if __name__ == '__main__':
    file_path = 'D:BaiduNetdiskDownload\a.mp3'
    file_path = mp3towav('D:BaiduNetdiskDownload\a.mp3', file_path.replace('mp3', 'wav'))
    data, _, _ = pre_deal(file_path)
    plpot(data)

• 通过wave库，可以识别音频文件，声道，样本宽度，帧率，帧数等
• 由于文件的左右声道的值都一样，所以简单处理只要其中一个声道
• 为了提升机器学习速度，修改采样率为20分之一，来降低数据量
• wave库只支持wav文件，所以需要把mp3转换为wav，这里用到了pydub库
• 解析音频数据后，通过matplotlib库来画图，显示出波纹图

二、人工标记数据

使用音频处理软件goldwave，采用人工听的方法来把音频文件的音乐部分剪掉，保存的文件放在chg目录里面，剪之前的文件放在raw目录下面。一共剪了18个文件。

三、获取训练数据

class LeaningTest():
    chg_path = r'D:BaiduNetdiskDownload	estchg'
    raw_path = r'D:BaiduNetdiskDownload	est
aw'
    model = None

    @classmethod
    def load_model(cls):
        cls.model = pickle_utils.load('knn.model.pkl')

    @classmethod
    def chg(cls):
        chg_path = r'D:BaiduNetdiskDownload	estchg'
        raw_path = r'D:BaiduNetdiskDownload	est
aw'
        for i, f in enumerate(os.listdir(chg_path)):
            shutil.copy(chg_path + '\' + f, chg_path + '\' + '%s.mp3' % i)
            shutil.copy(raw_path + '\' + f, raw_path + '\' + '%s.mp3' % i)

    @classmethod
    def get_path(cls, i, t):
        p = cls.chg_path if t == 'chg' else cls.raw_path
        return p + '\' + '%s.mp3' % i

    @classmethod
    def sample_cnt(cls, sample):
        """
        转换样本数据，返回每个区间的计数。
        例如从[0.1,0.1,0.8]转换为[2,1]
        2是[0,0.5)区间的计数
        1是[0.5,1)区间的计数
        """
        step = 0.025
        qujians = []
        start = 0
        while start < 1:
            qujians.append((start, start + step))
            start += step
        new_sample = [0 for i in range(len(qujians))]
        for s in sample:
            for i, qujian in enumerate(qujians):
                if qujian[0] <= s < qujian[1]:
                    new_sample[i] += 1
        return new_sample

    @classmethod
    def get_sample(cls, i):
        """
        获取用于机器学习的数据
        return [([100,200],0)]
        """
        chg = cls.to_wav(cls.get_path(i, 'chg'))
        raw = cls.to_wav(cls.get_path(i, 'raw'))

        data_chg, framerate_chg, n_frames_chg = pre_deal(chg)
        total_sec_chg = int(n_frames_chg / framerate_chg)

        data_raw, framerate_raw, n_frames_raw = pre_deal(raw)
        total_sec_raw = int(n_frames_raw / framerate_raw)

        length = 1
        samples = []
        for i in range(60, total_sec_raw, length):
            if total_sec_chg + 5 < i < total_sec_chg + 5:
                continue  # 不要这部分

            flag = 0 if i < total_sec_chg else 1
            # print get_index(framerate, 0, i),get_index(framerate, 0, i + length),total_sec
            sample = data_raw[get_index(framerate_raw, 0, i):get_index(framerate_raw, 0, i + length)]

            sample = cls.sample_cnt(sample)

            samples.append((sample, flag))
        return samples

    @classmethod
    def to_wav(cls, file_path):
        """转换mp3为wav"""
        if 'mp3' in file_path:
            to_file_path = file_path.replace('mp3', 'wav')
            mp3towav(file_path, to_file_path)
            file_path = to_file_path
        return file_path

    @classmethod
    def get_all_sample(cls, ):
        """获取所有样本"""
        file_name = 'sample4.json'
        if os.path.exists(file_name):
            with open(file_name, 'r') as f:
                return json.loads(f.read())
        else:
            samples = []
            for i in range(1):
                print 'get sample', i
                samples.extend(cls.get_sample(i))
            with open(file_name, 'w') as f:
                f.write(json.dumps(samples))
            return samples

    @classmethod
    def train_wrapper(cls):
        """训练"""
        samples = cls.get_all_sample()
        label0 = [s for s in samples if s[1] == 0]
        label1 = [s for s in samples if s[1] == 1]
        random.shuffle(label0)
        random.shuffle(label1)
        train_datas_sets = [i[0] for i in label0[:int(len(label0) * 0.7)]] + [i[0] for i in
                                                                              label1[:int(len(label1) * 0.7)]]
        train_labels_set = [i[1] for i in label0[:int(len(label0) * 0.7)]] + [i[1] for i in
                                                                              label1[:int(len(label1) * 0.7)]]
        test_datas_set = [i[0] for i in label0[int(len(label0) * 0.7):]] + [i[0] for i in
                                                                            label1[int(len(label1) * 0.7):]]
        test_labels_set = [i[1] for i in label0[int(len(label0) * 0.7):]] + [i[1] for i in
                                                                             label1[int(len(label1) * 0.7):]]
        print len(train_datas_sets)
        # cls.train_knn(train_datas_sets, train_labels_set, test_datas_set, test_labels_set)

 

if __name__ == '__main__':
    LeaningTest.train_wrapper()

• 以1秒钟为一个样本，然后对数据进行计数，返回每个区间的计数，区间间隔是0.025，所以一个样布的向量长度是40
• 由于前60s都是前奏，所以不作为训练数据
• 由于是人工分割，所以可能有误差，所以把分割点前后5s的都不作为训练数据

四、训练

@classmethod
def train(cls, train_datas_sets, train_labels_set, test_datas_set, test_labels_set):
    """
    """
    from sklearn.naive_bayes import GaussianNB
    from sklearn.linear_model import LogisticRegression
    from sklearn.linear_model import LinearRegression
    from sklearn import tree
    from sklearn import svm
    from sklearn.neural_network import MLPClassifier
    from sklearn import neighbors
    for mechine in [svm.SVC, LogisticRegression, LinearRegression, tree.DecisionTreeClassifier,
                    neighbors.KNeighborsClassifier, MLPClassifier, GaussianNB]:
        clf = mechine()
        clf.fit(train_datas_sets, train_labels_set)  # 训练
        score = clf.score(test_datas_set, test_labels_set)  # 预测测试集，并计算正确率
        print 'score', mechine, score

训练结果:

score <class 'sklearn.svm.classes.SVC'> 0.7203252032520325
score <class 'sklearn.linear_model.logistic.LogisticRegression'> 0.8886178861788618
score <class 'sklearn.linear_model.base.LinearRegression'> 0.40864632529611417
score <class 'sklearn.tree.tree.DecisionTreeClassifier'> 0.8888888888888888
score <class 'sklearn.neighbors.classification.KNeighborsClassifier'> 0.9224932249322493
score <class 'sklearn.neural_network.multilayer_perceptron.MLPClassifier'> 0.835230352303523
score <class 'sklearn.naive_bayes.GaussianNB'> 0.8035230352303523

• 使用多种模型进行训练，得到的结果为knn的准确率最高，达到了0.92

所以训练knn模型，并保存为pickle

@classmethod
def train_knn(cls, train_datas_sets, train_labels_set, test_datas_set, test_labels_set):
    from sklearn import neighbors
    mechine = neighbors.KNeighborsClassifier
    clf = mechine()
    clf.fit(train_datas_sets, train_labels_set)  # 训练
    score = clf.score(test_datas_set, test_labels_set)  # 预测测试集，并计算正确率
    print 'score', mechine, score
    pickle_utils.dump(clf, 'knn.model.pkl')

五、分割文件

@classmethod
def get_cut_sce(cls, file_path, model):
    """获取分割的秒数，找不到返回None"""
    file_path = cls.to_wav(file_path)
    data_raw, framerate, n_frames = pre_deal(file_path)
    total_sec = int(n_frames / framerate)

    length = 1
    rets = []
    for i in range(60, total_sec, length):
        # print file_path, i
        sample = data_raw[get_index(framerate, 0, i):get_index(framerate, 0, i + length)]

        sample = cls.sample_cnt(sample)
        ret = model.predict([sample])
        rets.append(ret)
        if ret == 1 and len(rets) >= 3 and rets[-2] == 1 and rets[-3] == 1:
            return i

    return None

@classmethod
def get_min(cls, sec):
    """转换秒数为 分秒格式"""
    print '%s:%s' % (int(sec / 60), int(sec % 60))

@classmethod
def predict(cls, ):
    """预测"""
    file_path = r'D:BaiduNetdiskDownloadc.mp3'
    model = pickle_utils.load('knn.model.pkl')
    sec = cls.get_cut_sce(file_path, model)
    print 'sec', sec, cls.get_min(sec)

@classmethod
def cut_song(cls, file_path, to_file_path, file_name):
    """分割歌曲"""
    print 'cut_song', file_name.decode('gbk'), file_path
    sec = cls.get_cut_sce(file_path, cls.model)
    if sec is None:
        print 'error can not find sec', file_path, file_name.decode('gbk')
        return 0
    song = AudioSegment.from_mp3(file_path)
    # to_file_path=file_path.replace('mp3','wav')
    song = song[:sec * 1000]
    song.export(to_file_path, 'mp3', bitrate='64k')
    return 1


@classmethod
def cut_songs(cls, ):
    """分割某个文件夹下面的所有歌曲"""
    root_path = r'D:BaiduNetdiskDownload听世界-战国5(156集)64kbps'
    del_path = r'D:BaiduNetdiskDownload	o_del'
    for f in os.listdir(root_path):
        if 'mp3' in f and 'cut' not in f:
            file_path = root_path + '\' + f
            if os.path.exists(file_path + '.cut.mp3'):
                print 'exist', file_path.decode('gbk') + '.cut.mp3'
                continue
            # 由于pydub不支持windows的中文路径，所以只能把源文件已到一个临时的英文目录，然后执行分割 然后把临时文件移走
            tmp_file_path = 'D:BaiduNetdiskDownload\test.mp3'  # pydub不支持中文地址，只能这样
            tmp_wav_path = tmp_file_path.replace('mp3', 'wav')
            tmp_to_file_path = tmp_file_path + '.cut.mp3'
            shutil.copy(file_path, tmp_file_path)
            ret = cls.cut_song(tmp_file_path, tmp_to_file_path,f)
            shutil.move(tmp_file_path, del_path + '\del1_' + f)
            shutil.move(tmp_wav_path, del_path + '\del3_' + f)
            try:
                # 有可能找不到分割点，导致没有分割，所以加上try
                shutil.copy(tmp_to_file_path, file_path + '.cut.mp3')
                shutil.move(tmp_to_file_path, del_path + '\del2_' + f)

            except:
                import traceback
                print traceback.format_exc()




@classmethod
def test(cls):
    song = AudioSegment.from_mp3(u'D:BaiduNetdiskDownload测试\a.mp3'.encode('gbk'))


if __name__ == '__main__':
    LeaningTest.load_model()
    LeaningTest.cut_songs()

• 即使准确率达到0.92，但是还没有到100%，所以连续3s都判断为音乐，才分割，这样理论的准确率可以去到1-0.08^3。
• 由于pydub不支持windows的中文路径，所以只能把源文件已到一个临时的英文目录，然后执行分割 然后把临时文件移走

未经同意，请不要转载