《自然語言處理技術(shù)》代碼42所示。代碼STYLEREF1\s4SEQ代碼\*ARABIC\s12對音頻數(shù)據(jù)進行MFCC特征提取defcompute_mfcc(file):#讀取音頻文件 fs,audio=wavfile.read(file)#使用mfcc函數(shù)計算音頻文件的特征 mfcc_feat=mfcc(audio,samplerate=(fs/2),numcep=26)#返回MFCC特征 returnmfcc_feat構(gòu)建GMM-HMM語音識別模型這一步驟，最重要的部分就是通過hmmlearn庫中的hmm模塊創(chuàng)建一個GMMHMM對象，并設(shè)置合適的參數(shù)，如n_components等，然后使用訓(xùn)練數(shù)據(jù)對模型進行訓(xùn)練，以學(xué)習(xí)不同音頻類別的模型參數(shù)，如REF_Ref98942649\h代碼43所示。代碼STYLEREF1\s4SEQ代碼\*ARABIC\s13構(gòu)建GMM-HMM語音識別模型#'''&usage: 搭建GMM-HMM的孤立詞識別模型參數(shù)意義： CATEGORY: 所有標(biāo)簽的列表 n_comp: 每個孤立詞中的狀態(tài)數(shù) cov_type: 協(xié)方差矩陣的類型 n_iter: 訓(xùn)練迭代次數(shù)'''#classModel():"""docstringforModel"""def__init__(self,CATEGORY=None,n_comp=3,n_mix=3,cov_type='diag',n_iter=1000):super(Model,self).__init__()self.CATEGORY=CATEGORYself.category=len(CATEGORY)self.n_comp=n_compself.n_mix=n_mixself.cov_type=cov_typeself.n_iter=n_iter#關(guān)鍵步驟，初始化models，返回特定參數(shù)的模型的列表self.models=[]forkinrange(self.category):model=hmm.GMMHMM(n_components=self.n_comp,n_mix=self.n_mix,covariance_type=self.cov_type,n_iter=self.n_iter)self.models.append(model)#模型訓(xùn)練deftrain(self,wavdict=None,labeldict=None):forkinrange(10):subdata=[]model=self.models[k]forxinwavdict:iflabeldict[x]==self.CATEGORY[k]:mfcc_feat=compute_mfcc(wavdict[x])model.fit(mfcc_feat)#使用特定的測試集合進行測試deftest(self,wavdict=None,labeldict=None):result=[]forkinrange(self.category):subre=[]label=[]model=self.models[k]forxinwavdict:mfcc_feat=compute_mfcc(wavdict[x])#生成每個數(shù)據(jù)在當(dāng)前模型下的得分情況re=model.score(mfcc_feat)subre.append(re)label.append(labeldict[x])#匯總得分情況result.append(subre)#選取得分最高的種類result=np.vstack(result).argmax(axis=0)#返回種類的類別標(biāo)簽result=[self.CATEGORY[label]forlabelinresult]print('識別得到結(jié)果：\n',result)print('原始標(biāo)簽類別：\n',label)#檢查識別率，為：正確識別的個數(shù)/總數(shù)totalnum=len(label)correctnum=0foriinrange(totalnum):ifresult[i]==label[i]:correctnum+=1print('識別率:',correctnum/totalnum)defsave(self,path="models.pkl"):#利用externaljoblib保存生成的hmm模型joblib.dump(self.models,path)defload(self,path="models.pkl"):#導(dǎo)入hmm模型self.models=joblib.load(path)GMM-HMM模型測試與訓(xùn)練這一步驟，使用測試數(shù)據(jù)對模型進行測試和評估，計算模型在測試集上的準(zhǔn)確率或其他評價指標(biāo)，最后也可以根據(jù)需要，可以進行模型參數(shù)調(diào)優(yōu)和訓(xùn)練過程的優(yōu)化，如REF_Ref98521257\h代碼44所示。代碼STYLEREF1\s4SEQ代碼\*ARABIC\s14GMM-HMM模型訓(xùn)練并測試訓(xùn)練集效果models=Model(CATEGORY=CATEGORY)models.train(wavdict=wavdict,labeldict=labeldict)models.save()models.load()models.test(wavdict=wavdict,labeldict=labeldict)運行REF_Ref98521257\h代碼44，得到訓(xùn)練集識別結(jié)果如下。識別得到結(jié)果：['4','4','6','4','5','5','10','10','4','4','10','4','4','10','10','10','7','2','10','6','8','4','4','2','4','4','2','4','10','2','9','10','8','10','7','10','8','10','10','7','10','10','4','4','8','6','10','10','7','10','1','4','4','4','10','4','6','5','10','10','4','2','4','7','8','4','6','2','7','4','1','10','8','3','10','8','4','6','5','4','1','10','9','10','10','9','8','8','10','4','7','9','10','9','6','5','8','7','10','10']原始標(biāo)簽類別：['4','3','6','4','5','5','1','1','4','6','9','4','6','9','1','9','7','2','1','6','8','4','5','2','3','4','2','6','1','2','10','10','8','9','7','9','8','7','1','7','1','2','3','3','8','5','10','2','7','2','1','3','4','3','5','3','6','5','10','10','3','2','3','7','8','2','5','2','7','4','1','5','8','3','9','8','6','6','5','4','1','10','9','6','10','9','8','8','9','4','7','7','10','9','6','5','8','7','10','10']識別率:0.6接下來對測試集機進行測試，如REF_Ref133925266\h代碼45所示。代碼STYLEREF1\s4SEQ代碼\*ARABIC\s15GMM-HMM模型測試集效果mod