acousticModelTraining.py 文件源码

def processAcousticModelTrainPho(class_name,syllableTierName,phonemeTierName,featureFilename,gmmModel_path):
    '''
    Monophonic acoustic model training
    :param mode: sourceSeparation, qmLonUpfLaosheng
    :param syllableTierName: 'pinyin', 'dian'
    :param phonemeTierName: 'details'
    :param featureFilename: 'dic_pho_feature_train.pkl'
    :param gmmModel_path: in parameters.py
    :return:
    '''

    # model training
    recordings_train = getRecordingNamesSimi('TRAIN',class_name)

    dic_pho_feature_train = dumpFeaturePho(class_name,recordings_train,syllableTierName,phonemeTierName)

    output = open(featureFilename, 'wb')
    pickle.dump(dic_pho_feature_train, output)
    output.close()

    # model loading
    pkl_file = open(featureFilename, 'rb')
    dic_pho_feature_train = pickle.load(pkl_file)
    pkl_file.close()

    n_comp = 40
    g = mixture.GaussianMixture(n_components=n_comp,covariance_type='diag')

    print len(dic_pho_feature_train.keys())
    for ii,key in enumerate(dic_pho_feature_train):
        # print key, dic_pho_feature_train[key].shape

        print 'fitting gmm ', key, ' ', str(ii), ' of ', str(len(dic_pho_feature_train.keys()))

        ##-- try just fit the first dim of MFCC
        # x = np.expand_dims(dic_pho_feature_train[key][:,0],axis=1)

        x = dic_pho_feature_train[key]
        print x.shape

        if x.shape[0] > n_comp*5:
            g.fit(x)

            output = open(os.path.join(gmmModel_path,key+'.pkl'),'wb')
            pickle.dump(g, output)
            output.close()
        else:
            # yn not fitted, because too few samples
            print(key+' is not fitted.')