forked from shunliz/Machine-Learning
-
Notifications
You must be signed in to change notification settings - Fork 0
/
gmm-hmm.md
473 lines (125 loc) · 16.2 KB
/
gmm-hmm.md
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
本文简明讲述GMM-HMM在语音识别上的原理,建模和测试过程。这篇blog只回答三个问题:
1. 什么是[Hidden Markov Model](http://en.wikipedia.org/wiki/Hidden_Markov_models)?
HMM要解决的三个问题:
1\) Likelihood
2\) Decoding
3\) Training
2. GMM是神马?怎样用GMM求某一音素(phoneme)的概率?
3. GMM+HMM大法解决语音识别
3.1 识别
3.2 训练
3.2.1 Training the params of GMM
3.2.2 Training the params of HMM
首先声明我是做视觉的不是做语音的,迫于\*\*需要24小时速成语音。上网查GMM-HMM资料中文几乎为零,英文也大多是paper。苦苦追寻终于貌似搞懂了GMM-HMM,感谢语音组老夏([http://weibo.com/ibillxia](http://weibo.com/ibillxia))提供资料给予指导。本文结合最简明的概括还有自己一些理解应运而生,如有错误望批评指正。
====================================================================
1. 什么是[Hidden Markov Model](http://en.wikipedia.org/wiki/Hidden_Markov_models)?
![](http://img.blog.csdn.net/20140528174242250?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
ANS:一个有隐节点(unobservable)和可见节点(visible)的马尔科夫过程(见[详解](http://blog.csdn.net/abcjennifer/article/details/25908495))。
隐节点表示状态,可见节点表示我们听到的语音或者看到的时序信号。
最开始时,我们指定这个HMM的结构,训练HMM模型时:给定n个时序信号y1...yT(训练样本), 用MLE(typicallyimplemented in EM) 估计参数:
1. N个状态的初始概率
2. 状态转移概率a
3. 输出概率b
--------------
* 在语音处理中,一个word由若干phoneme(音素)组成;
* 每个HMM对应于一个word或者音素(phoneme)
* 一个word表示成若干states,每个state表示为一个音素
用HMM需要解决3个问题:
1).Likelihood: 一个HMM生成一串observation序列x的概率< the Forward algorithm>
![](http://img.blog.csdn.net/20140530151854546?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
其中,αt(sj)表示HMM在时刻t处于状态j,且observation = {x1,...,xt}的概率![](http://img.blog.csdn.net/20140530152949593?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast),
aij是状态i到状态j的转移概率,
bj(xt)表示在状态j的时候生成xt的概率,
2).Decoding: 给定一串observation序列x,找出最可能从属的HMM状态序列< the Viterbi algorithm>
在实际计算中会做剪枝,不是计算每个可能state序列的probability,而是用Viterbi approximation:
从时刻1:t,只记录转移概率最大的state和概率。
记Vt\(si\)为从时刻t-1的所有状态转移到时刻t时状态为j的最大概率:![](http://img.blog.csdn.net/20140530155625171?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
记![](http://img.blog.csdn.net/20140530154949078?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)为:从时刻t-1的哪个状态转移到时刻t时状态为j的概率最大;
进行Viterbi approximation过程如下:
![](http://img.blog.csdn.net/20140530155945437?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
然后根据记录的最可能转移状态序列![](http://img.blog.csdn.net/20140530154949078?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)进行回溯:
![](http://img.blog.csdn.net/20140530160136578?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
3).Training: 给定一个observation序列x,训练出HMM参数λ = {aij, bij} the EM \(Forward-Backward\) algorithm
这部分我们放到“3. GMM+HMM大法解决语音识别”中和GMM的training一起讲
---------------------------------------------------------------------
2. GMM是神马?怎样用GMM求某一音素(phoneme)的概率?
2.1 简单理解混合高斯模型就是几个高斯的叠加。。。e.g. k=3
![](http://img.blog.csdn.net/20140528180736578?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![](http://img.blog.csdn.net/20140530134729015?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig2. GMM illustration and the probability of x
2.2 GMM for state sequence
每个state有一个GMM,包含k个高斯模型参数。如”hi“(k=3):
PS:sil表示silence(静音)
![](http://img.blog.csdn.net/20140528200425421?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig3. use GMM to estimate the probability of a state sequence given observation {o1, o2, o3}
其中,每个GMM有一些参数,就是我们要train的输出概率参数
![](http://img.blog.csdn.net/20140528200531906?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig4. parameters of a GMM
怎么求呢?和KMeans类似,如果已知每个点x^n属于某每类 j 的概率p\(j\|x^n\),则可以估计其参数:
![](http://img.blog.csdn.net/20140530135251546?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast) , 其中 ![](http://img.blog.csdn.net/20140530135311953?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
只要已知了这些参数,我们就可以在predict(识别)时在给定input sequence的情况下,计算出一串状态转移的概率。如上图要计算的state sequence 1->2->2概率:
![](http://img.blog.csdn.net/20140528201041078?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig5. probability of S1->S2->S3 given o1->o2->o3
---------------------------------------------------------------------
3. GMM+HMM大法解决语音识别
<!--识别-->
我们获得observation是语音waveform, 以下是一个词识别全过程:
1\). 将waveform切成等长frames,对每个frame提取特征(e.g. MFCC),
2\).对每个frame的特征跑GMM,得到每个frame\(o\_i\)属于每个状态的概率b\_state\(o\_i\)
![](http://img.blog.csdn.net/20140528203714828?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig6. complete process from speech frames to a state sequence
3\). 根据每个单词的HMM状态转移概率a计算每个状态sequence生成该frame的概率; 哪个词的HMM 序列跑出来概率最大,就判断这段语音属于该词
宏观图:
![](http://img.blog.csdn.net/20140528175313171?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig7. Speech recognition, a big framework
\(from Encyclopedia of Information Systems, 2002\)
<!--训练-->
好了,上面说了怎么做识别。那么我们怎样训练这个模型以得到每个GMM的参数和HMM的转移概率什么的呢?
①Training the params of GMM
GMM参数:高斯分布参数:![](http://img.blog.csdn.net/20140530185018734?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
从上面fig4下面的公式我们已经可以看出来想求参数必须要知道P\(j\|x\),即,x属于第j个高斯的概率。怎么求捏?
![](http://img.blog.csdn.net/20140530141637656?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
fig8. bayesian formula of P\( j \| x \)
根据上图 P(j \| x), 我们需要求P\(x\|j\)和P(j)去估计P\(j\|x\).
这里由于P\(x\|j\)和P(j)都不知道,需要用EM算法迭代估计以最大化P\(x\) = P\(x1\)\*p\(x2\)\*...\*P\(xn\):
A. 初始化(可以用kmeans)得到P\(j\)
B. 迭代
E(estimate)-step: 根据当前参数 \(means, variances, mixing parameters\)估计P\(j\|x\)
M(maximization)-step: 根据当前P\(j\|x\) 计算GMM参数(根据fig4 下面的公式:)
![](http://img.blog.csdn.net/20140530135251546?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast) , 其中 ![](http://img.blog.csdn.net/20140530135311953?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
②Training the params of HMM
前面已经有了GMM的training过程。在这一步,我们的目标是:从observation序列中估计HMM参数λ;
假设状态->observation服从单核高斯概率分布:![](http://img.blog.csdn.net/20140530162550421?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast),则λ由两部分组成:
![](http://img.blog.csdn.net/20140530195145953?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
HMM训练过程:迭代
E(estimate)-step: 给定observation序列,估计时刻t处于状态sj的概率 ![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
M(maximization)-step: 根据![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)重新估计HMM参数aij.
其中,
E-step: 给定observation序列,估计时刻t处于状态sj的概率 ![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
为了估计![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast), 定义![](http://img.blog.csdn.net/20140530191032625?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast): t时刻处于状态sj的话,t时刻未来observation的概率。即![](http://img.blog.csdn.net/20140530191206000?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
这个可以递归计算:β\_t(si)=从状态 si 转移到其他状态 sj 的概率aij\*状态 i 下观测到x\_{t+1}的概率bi\(x\_{t+1}\)\*t时刻处于状态sj的话{t+1}后observation概率β\_{t+1}(sj)
即:
![](http://img.blog.csdn.net/20140530191353765?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
定义刚才的![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)为state occupation probability,表示给定observation序列,时刻t处于状态sj的概率P\(S\(t\)=sj \| X,λ\)。根据贝叶斯公式p\(A\|B,C\) = P\(A,B\|C\)/P\(B\|C\),有:
![](http://img.blog.csdn.net/20140530194138937?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
由于分子p\(A,B\|C\)为
![](http://img.blog.csdn.net/20140530193757812?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
其中,αt(sj)表示HMM在时刻t处于状态j,且observation = {x1,...,xt}的概率![](http://img.blog.csdn.net/20140530152949593?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast);
![](http://img.blog.csdn.net/20140530191032625?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast): t时刻处于状态sj的话,t时刻未来observation的概率;
且![](http://img.blog.csdn.net/20140530193617734?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
finally, 带入![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)的定义式有:
![](http://img.blog.csdn.net/20140530194816484?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
好,终于搞定!对应上面的E-step目标,只要给定了observation和当前HMM参数 λ,我们就可以估计![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)了对吧 \(\*^\_\_^\*\)
M-step:根据![](http://img.blog.csdn.net/20140530185647156?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)重新估计HMM参数λ:
对于λ中高斯参数部分,和GMM的M-step是一样一样的(只不过这里写成向量形式):
![](http://img.blog.csdn.net/20140530200004781?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
对于λ中的状态转移概率aij, 定义C\(Si->Sj\)为从状态Si转到Sj的次数,有
![](http://img.blog.csdn.net/20140530200136921?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
实际计算时,定义每一时刻的转移概率![](http://img.blog.csdn.net/20140530200404828?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)为时刻t从si->sj的概率:
![](http://img.blog.csdn.net/20140530200424640?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
那么就有:
![](http://img.blog.csdn.net/20140530200615750?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
把HMM的EM迭代过程和要求的参数写专业点,就是这样的:
![](http://img.blog.csdn.net/20140530200730218?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvYWJjamVubmlmZXI=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
PS:这个训练HMM的算法叫 Forward-Backward algorithm。
一个很好的reference:[点击打开链接](http://www.inf.ed.ac.uk/teaching/courses/asr/2012-13/asr03-hmmgmm-4up.pdf)