本文，我们将创建一个简单的python机器学习算法，以便能够通过声音来诊断该人是否为患者。
我们将使用一组(健康者和帕金森病患者)音频文件库，通过对音频进行一些测量来构建我们的机器学习数据集。建立机器学习数据集后，我们将使用SciKit Learn库训练线性回归模型。最后，我们将构建一个python库，这个库可以轻松集成到其他应用程序中。
数据集
首先，我们需要将音频文件转换成包含音频测量值以及患者是否健康标识的表格。
我们将要使用的音频文件(https://zenodo.org/record/2867216#.Xp4kVsgzaUl )。
让我们从导入必要的Python库开始。
import globimport numpy as npimport pandas as pdimport parselmouthfrom parselmouth.praat import call
接下来，我们将创建一个函数，该函数允许您对输入音频文件进行各种复杂的测量。这些测量是通过parselmouth库实现的，它允许在python代码中使用praat。(https://parselmouth.readthedocs.io/en/stable/)
def measurePitch(voiceID, f0min, f0max, unit): sound = parselmouth.Sound(voiceID) # read the sound pitch = call(sound, "To Pitch", 0.0, f0min, f0max) pointProcess = call(sound, "To PointProcess (periodic, cc)", f0min, f0max)#create a praat pitch object localJitter = call(pointProcess, "Get jitter (local)", 0, 0, 0.0001, 0.02, 1.3) localabsoluteJitter = call(pointProcess, "Get jitter (local, absolute)", 0, 0, 0.0001, 0.02, 1.3) rapJitter = call(pointProcess, "Get jitter (rap)", 0, 0, 0.0001, 0.02, 1.3) ppq5Jitter = call(pointProcess, "Get jitter (ppq5)", 0, 0, 0.0001, 0.02, 1.3) localShimmer = call([sound, pointProcess], "Get shimmer (local)", 0, 0, 0.0001, 0.02, 1.3, 1.6) localdbShimmer = call([sound, pointProcess], "Get shimmer (local_dB)", 0, 0, 0.0001, 0.02, 1.3, 1.6) apq3Shimmer = call([sound, pointProcess], "Get shimmer (apq3)", 0, 0, 0.0001, 0.02, 1.3, 1.6) aqpq5Shimmer = call([sound, pointProcess], "Get shimmer (apq5)", 0, 0, 0.0001, 0.02, 1.3, 1.6) apq11Shimmer = call([sound, pointProcess], "Get shimmer (apq11)", 0, 0, 0.0001, 0.02, 1.3, 1.6) harmonicity05 = call(sound, "To Harmonicity (cc)", 0.01, 500, 0.1, 1.0) hnr05 = call(harmonicity05, "Get mean", 0, 0) harmonicity15 = call(sound, "To Harmonicity (cc)", 0.01, 1500, 0.1, 1.0) hnr15 = call(harmonicity15, "Get mean", 0, 0) harmonicity25 = call(sound, "To Harmonicity (cc)", 0.01, 2500, 0.1, 1.0) hnr25 = call(harmonicity25, "Get mean", 0, 0) harmonicity35 = call(sound, "To Harmonicity (cc)", 0.01, 3500, 0.1, 1.0) hnr35 = call(harmonicity35, "Get mean", 0, 0) harmonicity38 = call(sound, "To Harmonicity (cc)", 0.01, 3800, 0.1, 1.0) hnr38 = call(harmonicity38, "Get mean", 0, 0) return localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38
然后，我们为每种类型的测量创建一个列表，再创建一个列表用来表示病人是否健康。在列表被填入后用来构建机器学习数据集。
localJitter_list = [] #measurelocalabsoluteJitter_list = [] #measurerapJitter_list = [] #measureppq5Jitter_list = [] #measurelocalShimmer_list = [] #measurelocaldbShimmer_list = [] #measureapq3Shimmer_list = [] #measureaqpq5Shimmer_list = [] #measureapq11Shimmer_list = [] #measurehnr05_list = [] #measurehnr15_list = [] #measurehnr25_list = [] #measureparkinson_list = [] #Parkinson(1) or healthy(0)
现在，我们可以使用前面创建的函数通过对音频文件进行测量来填充列表。我们需要使用4个for循环来填充列表。
for wave_file in glob.glob("audio/SpontaneousDialogue/PD/*.wav"): sound = parselmouth.Sound(wave_file) (localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38) = measurePitch(sound, 75, 1000, "Hertz") file_list.append(wave_file) # make an ID list localJitter_list.append(localJitter) # make a mean F0 list localabsoluteJitter_list.append(localabsoluteJitter) # make a sd F0 list rapJitter_list.append(rapJitter) ppq5Jitter_list.append(ppq5Jitter) localShimmer_list.append(localShimmer) localdbShimmer_list.append(localdbShimmer) apq3Shimmer_list.append(apq3Shimmer) aqpq5Shimmer_list.append(aqpq5Shimmer) apq11Shimmer_list.append(apq11Shimmer) hnr05_list.append(hnr05) hnr15_list.append(hnr15) hnr25_list.append(hnr25) parkinson_list.append(1) #1 because parkinson filefor wave_file in glob.glob("audio/ReadText/PD/*.wav"): sound = parselmouth.Sound(wave_file) (localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38) = measurePitch(sound, 75, 1000, "Hertz") file_list.append(wave_file) # make an ID list localJitter_list.append(localJitter) # make a mean F0 list localabsoluteJitter_list.append(localabsoluteJitter) # make a sd F0 list rapJitter_list.append(rapJitter) ppq5Jitter_list.append(ppq5Jitter) localShimmer_list.append(localShimmer) localdbShimmer_list.append(localdbShimmer) apq3Shimmer_list.append(apq3Shimmer) aqpq5Shimmer_list.append(aqpq5Shimmer) apq11Shimmer_list.append(apq11Shimmer) hnr05_list.append(hnr05) hnr15_list.append(hnr15) hnr25_list.append(hnr25) parkinson_list.append(1) #1 because parkinson filefor wave_file in glob.glob("audio/SpontaneousDialogue/HC/*.wav"): sound = parselmouth.Sound(wave_file) (localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38) = measurePitch(sound, 75, 1000, "Hertz") file_list.append(wave_file) # make an ID list localJitter_list.append(localJitter) # make a mean F0 list localabsoluteJitter_list.append(localabsoluteJitter) # make a sd F0 list rapJitter_list.append(rapJitter) ppq5Jitter_list.append(ppq5Jitter) localShimmer_list.append(localShimmer) localdbShimmer_list.append(localdbShimmer) apq3Shimmer_list.append(apq3Shimmer) aqpq5Shimmer_list.append(aqpq5Shimmer) apq11Shimmer_list.append(apq11Shimmer) hnr05_list.append(hnr05) hnr15_list.append(hnr15) hnr25_list.append(hnr25) parkinson_list.append(0) #0 because healthy filefor wave_file in glob.glob("audio/ReadText/HC/*.wav"): sound = parselmouth.Sound(wave_file) (localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38) = measurePitch(sound, 75, 1000, "Hertz") file_list.append(wave_file) # make an ID list localJitter_list.append(localJitter) # make a mean F0 list localabsoluteJitter_list.append(localabsoluteJitter) # make a sd F0 list rapJitter_list.append(rapJitter) ppq5Jitter_list.append(ppq5Jitter) localShimmer_list.append(localShimmer) localdbShimmer_list.append(localdbShimmer) apq3Shimmer_list.append(apq3Shimmer) aqpq5Shimmer_list.append(aqpq5Shimmer) apq11Shimmer_list.append(apq11Shimmer) hnr05_list.append(hnr05) hnr15_list.append(hnr15) hnr25_list.append(hnr25) parkinson_list.append(0) #0 because healthy file
最后，借助于panda和numpy库，我们必须将这些列表分组到一个表中，从而将它们转换为机器学习数据集。
pred = pd.DataFrame(np.column_stack([parkinson_list,localJitter_list, localabsoluteJitter_list, rapJitter_list, ppq5Jitter_list, localShimmer_list, localdbShimmer_list, apq3Shimmer_list, aqpq5Shimmer_list, apq11Shimmer_list, hnr05_list, hnr15_list, hnr25_list]), columns=["Parkinson","Jitter_rel","Jitter_abs","Jitter_RAP","Jitter_PPQ","Shim_loc","Shim_dB","Shim_APQ3","Shim_APQ5","Shi_APQ11", "hnr05", "hnr15", "hnr25"]) #add these lists to pandas in the right orderpred['hnr25'].fillna((parkinson['hnr25'].mean()), inplace=True) #Data cleaning because they may be NaN valuespred['hnr15'].fillna((parkinson['hnr15'].mean()), inplace=True) #Data cleaning because they may be NaN valuespred.to_csv("processed_results.csv", index=False) # Write out the updated dataset
制作机器学习模型
我们将使用前面提到的SciKit learn库的线性回归算法，该算法允许我们根据几个参数(measures)对标签(0或1)进行分类。
首先，我们将通过指定参数(measurements)和标签(0或1)来训练我们的机器学习模型。
parkinson = pd.read_csv("processed_results.csv") #Loading CSV datasetpredictors=["Jitter_rel","Jitter_abs","Jitter_RAP","Jitter_PPQ","Shim_loc","Shim_dB","Shim_APQ3","Shim_APQ5","Shi_APQ11","hnr05","hnr15", "hnr25"] #Listing predictorsfor col in predictors: # Loop through all columns in predictors if parkinson[col].dtype == 'object': # check if column's type is object (text) parkinson[col] = pd.Categorical(parkinson[col]).codes # convert text to numerical from sklearn.model_selection import train_test_splitX_train, X_test, y_train, y_test = train_test_split(parkinson[predictors], parkinson['Parkinson'], test_size=0.25, random_state=1) from sklearn.linear_model import LogisticRegressionclf = LogisticRegression()clf.fit(X_train, y_train)train_score = clf.score(X_train, y_train)test_score = clf.score(X_test, y_test)print ('train accuracy =', train_score)print ('test accuracy =', test_score)#train accuracy = 0.6666666666666666#test accuracy = 0.631578947368421
我们获得0.63的精度，考虑数据集的数据如此有限，这个结果还是令人满意的。
导出机器学习模型的Python代码如下：
import joblibclf.fit(X_train, y_train)joblib.dump(clf, "trainedModel.sav")
制作库
请记住，我们的目标是获得一个可以被另一个程序使用的库。Python实现代码如下：
import joblibimport parselmouthfrom parselmouth.praat import callimport pandas as pdimport numpy as npimport sklearndef loadModel(PATH): clf = joblib.load(PATH) return clfdef measurePitch(voiceID, f0min, f0max, unit): sound = parselmouth.Sound(voiceID) # read the sound pitch = call(sound, "To Pitch", 0.0, f0min, f0max) pointProcess = call(sound, "To PointProcess (periodic, cc)", f0min, f0max)#create a praat pitch object localJitter = call(pointProcess, "Get jitter (local)", 0, 0, 0.0001, 0.02, 1.3) localabsoluteJitter = call(pointProcess, "Get jitter (local, absolute)", 0, 0, 0.0001, 0.02, 1.3) rapJitter = call(pointProcess, "Get jitter (rap)", 0, 0, 0.0001, 0.02, 1.3) ppq5Jitter = call(pointProcess, "Get jitter (ppq5)", 0, 0, 0.0001, 0.02, 1.3) localShimmer = call([sound, pointProcess], "Get shimmer (local)", 0, 0, 0.0001, 0.02, 1.3, 1.6) localdbShimmer = call([sound, pointProcess], "Get shimmer (local_dB)", 0, 0, 0.0001, 0.02, 1.3, 1.6) apq3Shimmer = call([sound, pointProcess], "Get shimmer (apq3)", 0, 0, 0.0001, 0.02, 1.3, 1.6) aqpq5Shimmer = call([sound, pointProcess], "Get shimmer (apq5)", 0, 0, 0.0001, 0.02, 1.3, 1.6) apq11Shimmer = call([sound, pointProcess], "Get shimmer (apq11)", 0, 0, 0.0001, 0.02, 1.3, 1.6) harmonicity05 = call(sound, "To Harmonicity (cc)", 0.01, 500, 0.1, 1.0) hnr05 = call(harmonicity05, "Get mean", 0, 0) harmonicity15 = call(sound, "To Harmonicity (cc)", 0.01, 1500, 0.1, 1.0) hnr15 = call(harmonicity15, "Get mean", 0, 0) harmonicity25 = call(sound, "To Harmonicity (cc)", 0.01, 2500, 0.1, 1.0) hnr25 = call(harmonicity25, "Get mean", 0, 0) harmonicity35 = call(sound, "To Harmonicity (cc)", 0.01, 3500, 0.1, 1.0) hnr35 = call(harmonicity35, "Get mean", 0, 0) harmonicity38 = call(sound, "To Harmonicity (cc)", 0.01, 3800, 0.1, 1.0) hnr38 = call(harmonicity38, "Get mean", 0, 0) return localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15 ,hnr25 ,hnr35 ,hnr38def predict(clf, wavPath): file_list = [] localJitter_list = [] localabsoluteJitter_list = [] rapJitter_list = [] ppq5Jitter_list = [] localShimmer_list = [] localdbShimmer_list = [] apq3Shimmer_list = [] aqpq5Shimmer_list = [] apq11Shimmer_list = [] hnr05_list = [] hnr15_list = [] hnr25_list = [] hnr35_list = [] hnr38_list = [] sound = parselmouth.Sound(wavPath) (localJitter, localabsoluteJitter, rapJitter, ppq5Jitter, localShimmer, localdbShimmer, apq3Shimmer, aqpq5Shimmer, apq11Shimmer, hnr05, hnr15, hnr25, hnr35, hnr38) = measurePitch(sound, 75, 1000, "Hertz") localJitter_list.append(localJitter) # make a mean F0 list localabsoluteJitter_list.append(localabsoluteJitter) # make a sd F0 list rapJitter_list.append(rapJitter) ppq5Jitter_list.append(ppq5Jitter) localShimmer_list.append(localShimmer) localdbShimmer_list.append(localdbShimmer) apq3Shimmer_list.append(apq3Shimmer) aqpq5Shimmer_list.append(aqpq5Shimmer) apq11Shimmer_list.append(apq11Shimmer) hnr05_list.append(hnr05) hnr15_list.append(hnr15) hnr25_list.append(hnr25) hnr35_list.append(hnr35) hnr38_list.append(hnr38) toPred = pd.DataFrame(np.column_stack( [localJitter_list, localabsoluteJitter_list, rapJitter_list, ppq5Jitter_list, localShimmer_list, localdbShimmer_list, apq3Shimmer_list, aqpq5Shimmer_list, apq11Shimmer_list, hnr05_list, hnr15_list, hnr25_list]), columns=["Jitter_rel", "Jitter_abs", "Jitter_RAP", "Jitter_PPQ", "Shim_loc", "Shim_dB", "Shim_APQ3", "Shim_APQ5", "Shi_APQ11", "hnr05", "hnr15", "hnr25"]) # add these lists to pandas in the right order resp = clf.predict(toPred) resp = str(resp) if resp == "[1.]": return True else: return False
调用上述库的Python代码如下：
from RecognitionLib import *path = "../trainedModel.sav" #Model pathclf = loadModel(path) #Model loadingprint(predict(clf, "../../audio/ok.wav"))#Predicition
最后
可以通过拥有更大的数据集来改进机器学习模型，从而获得更高的精度(即通过从帕金森氏病患者那里获得更多的音频样本)。