基於hmm-gmm的1到10喚醒-python

可供參考的項目有：

https://github.com/jayaram1125/Single-Word-Speech-Recognition-using-GMM-HMM-

 

 

數據集得自己造：

 

# -----------------------------------------------------------------------------------------------------
'''
&usage:		HMM-GMM的孤立詞識別模型
@author:	hongwen sun
'''
# -----------------------------------------------------------------------------------------------------
# 導入依賴庫，特別需要注意hmmlearn
from python_speech_features import mfcc
from scipy.io import wavfile
from hmmlearn import hmm
from sklearn.externals import joblib
import numpy as np
import os


# -----------------------------------------------------------------------------------------------------
'''
&usage:		準備所需數據
'''
# -----------------------------------------------------------------------------------------------------
# 生成wavdict，key=wavid，value=wavfile
def gen_wavlist(wavpath):
	wavdict = {}
	labeldict = {}
	for (dirpath, dirnames, filenames) in os.walk(wavpath):
		for filename in filenames:
			if filename.endswith('.wav'):
				filepath = os.sep.join([dirpath, filename])
				fileid = filename.strip('.wav')
				wavdict[fileid] = filepath
				label = fileid.split('_')[1]
				labeldict[fileid] = label
	return wavdict, labeldict




# 特徵提取，feat = compute_mfcc(wadict[wavid])
def compute_mfcc(file):
	fs, audio = wavfile.read(file)
	# 這裏我故意fs/2,有些類似減小step，不建議這樣做，投機取巧做法
	mfcc_feat = mfcc(audio, samplerate=(fs/2), numcep=26)
	return mfcc_feat


# -----------------------------------------------------------------------------------------------------
'''
&usage:		搭建HMM-GMM的孤立詞識別模型
參數意義：
	CATEGORY:	所有標籤的列表
	n_comp:		每個孤立詞中的狀態數
	n_mix:		每個狀態包含的混合高斯數量
	cov_type:	協方差矩陣的類型
	n_iter:		訓練迭代次數
'''
# -----------------------------------------------------------------------------------------------------

class Model():
	"""docstring for Model"""
	def __init__(self, CATEGORY=None, n_comp=3, n_mix = 3, cov_type='diag', n_iter=1000):
		super(Model, self).__init__()
		self.CATEGORY = CATEGORY
		self.category = len(CATEGORY)
		self.n_comp = n_comp
		self.n_mix = n_mix
		self.cov_type = cov_type
		self.n_iter = n_iter
		# 關鍵步驟，初始化models，返回特定參數的模型的列表
		self.models = []
		for k in range(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)

	# 模型訓練
	def train(self, wavdict=None, labeldict=None):
		for k in range(10):
			subdata = []
			model = self.models[k]
			for x in wavdict:
				if labeldict[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
	def test(self, wavdict=None, labeldict=None):
		result = []
		for k in range(self.category):
			subre = []
			label = []
			model = self.models[k]
			for x in wavdict:
				mfcc_feat = compute_mfcc(wavdict[x])
				# 生成每個數據在當前模型下的得分情況
				re = model.score(mfcc_feat)
				subre.append(re)
				label.append(labeldict[x])
			# 彙總得分情況
			result.append(subre)
		# 選取得分最高的種類
		result = np.vstack(result).argmax(axis=0)
		# 返回種類的類別標籤
		result = [self.CATEGORY[label] for label in result]
		print('識別得到結果：\n',result)
		print('原始標籤類別：\n',label)
		# 檢查識別率，爲：正確識別的個數/總數
		totalnum = len(label)
		correctnum = 0
		for i in range(totalnum):
		 	if result[i] == label[i]:
		 	 	correctnum += 1 
		print('識別率:', correctnum/totalnum)


	def save(self, path="models.pkl"):
		# 利用external joblib保存生成的hmm模型
		joblib.dump(self.models, path)


	def load(self, path="models.pkl"):
		# 導入hmm模型
		self.models = joblib.load(path)


# -----------------------------------------------------------------------------------------------------
'''
&usage:		使用模型進行訓練和識別
'''
# -----------------------------------------------------------------------------------------------------
# 準備訓練所需數據
CATEGORY = ['1', '2', '3', '4', '5', '6', '7', '8', '9', '10']
wavdict, labeldict = gen_wavlist('training_data')
testdict, testlabel = gen_wavlist('test_data')	
# 進行訓練
models = Model(CATEGORY=CATEGORY)
models.train(wavdict=wavdict, labeldict=labeldict)
models.save()
models.load()
models.test(wavdict=wavdict, labeldict=labeldict)
models.test(wavdict=testdict, labeldict=testlabel)