python sklearn 语音识别

文件下载地址:
链接: https://pan.baidu.com/s/1lJfh1HyEQ2dRSbAhVoJNSA 提取码: dj6q

"""
    语音识别
"""
import numpy as np
import scipy.io.wavfile as wf
import python_speech_features as sf
import sklearn.svm as svm
import sklearn.metrics as sm
import sklearn.preprocessing as sp
import os


# 整理样本
def search_files(directory):
    """
        检索目录下的所有wav文件 返回目录字典
        {“appple”:[url, url...],
        “kiwi”:[url, url...],....}
    """
    files_dict = {}
    for cur_dir, sub_dirs, files in os.walk(directory):
        for file in files:
            if file.endswith(".wav"):
                label = cur_dir.split(os.path.sep)[-1]
                if label not in files_dict:
                    files_dict[label] = []
                url = os.path.join(cur_dir, file)
                files_dict[label].append(url)
    return files_dict


def files_mfc(file_urls):
    """
        读取语音文件 并且 MFCC化音频文件 梅尔频率倒谱系数
        返回整理好的 样本数据
    """
    x_data, y_data = [], []
    for label, urls in file_urls.items():
        for file in urls:
            sample_rate, signs = wf.read(file)
            mfc = sf.mfcc(signs, sample_rate)
            # ==>(1,13)
            sample = np.mean(mfc, axis=0)  # 求平均值
            x_data.append(sample)   # 音频特征数组
            y_data.append(label)    # 标签特征数组
    x_data = np.array(x_data)
    return x_data, y_data


# 读取训练集 文件路径
train_urls = search_files("./speeches/training")
print(train_urls)

# 整理训练集
train_x, train_y = files_mfc(train_urls)
# 标签编码
encoder = sp.LabelEncoder()
train_y_label = encoder.fit_transform(train_y)
print(train_x.shape, train_y_label.shape)

# 创建模型 SVC 并训练
model = svm.SVC(kernel="poly", degree=2, gamma="auto", probability=True)
model.fit(train_x, train_y_label)

# 读取测试集 文件路径
test_urls = search_files("./speeches/testing")
print(test_urls)

# 整理测试集
test_x, test_y = files_mfc(test_urls)
test_y_label = encoder.transform(test_y)

# 预测测试集 输出分类报告
prd_test_y = model.predict(test_x)
print(sm.classification_report(test_y_label, prd_test_y))

# 输出置信概率
probs = model.predict_proba(test_x)
print(np.round(probs, 3))

# 循环打印结果
for label, prob in zip(encoder.inverse_transform(prd_test_y), probs.max(axis=1)):
    print(label, np.round(prob, 3))