HTK Chapter 3 - Section 1 - Step 2

Below paragraphs are belong to 

• HTKBooks, 
  
• 苏统华, 哈尔滨工业大学人工智能研究室, 2006年10月30日,
  
• Howard Hung-Ju Chou, Intelligence Information Retrieval Lab., NCKU, Taiwan(R.O.C.).
  

Environment: 

• HTK 3.4 
  
• Cygwin NT-5.1 1.5.25

Section 1 is Data Preparation - 資料準備

• Step 1 - the Task Grammer 辨識模型要用的文法資料(gram->wdnet)
  
• Step 2 - the Dictionary  辨識模型要用的字典資料 (利用字典將wlist裡的單字翻譯為phones字串, one -> w ah n)

Suppose we have the experiment data EXP like following,

================================================

S0001 ONE VALIDATED ACTS OF SCHOOL DISTRICTS

S0002 TWO OTHER CASES ALSO WERE UNDER ADVISEMENT

S0003 BOTH FIGURES WOULD GO HIGHER IN LATER YEARS

S0004 THIS IS NOT A PROGRAM OF SOCIALIZED MEDICINE

================================================

we hope to transform the sentence prompts to be sequence of phones.

For example, 

One -> w ah n

Validated -> v ae l ih d....

by the collection of phones, so called dictionary here.

Before we get the dictionary, we have to transform our experiment data EXP to wlist file.

In pp.26 of HTKBook, it mention that we can use perl cript, called "prompts2wlist.pl".

----------------------------------

$perl  prompts2wlist.pl EXP  wlist-test

----------------------------------

Then you will get this wlist-test file, words are sorted.

=======================

A

ACTS

ADVISEMENT

ALSO

BOTH

CASES

DISTRICTS

FIGURES

GO

HIGHER

IN

IS

LATER

MEDICINE

NOT

OF

ONE

OTHER

PROGRAM

SCHOOL

SOCIALIZED

THIS

TWO

UNDER

VALIDATED

WERE

WOULD

YEARS

======================

It is parsed by the four sentence in EXP file,

========================

S0001 ONE VALIDATED ACTS OF SCHOOL DISTRICTS

S0002 TWO OTHER CASES ALSO WERE UNDER ADVISEMENT

S0003 BOTH FIGURES WOULD GO HIGHER IN LATER YEARS

S0004 THIS IS NOT A PROGRAM OF SOCIALIZED MEDICINE

========================

Then, we can use HDMan Dictionary MAnager to constrcut the new files by referring dictionaries.

HDMan can mrege different dictionaries and transcript the words in inputfile to the phones description.

---------------------------------------------------------------------------------------------------

$ HDMan -m  -w input_word_list -n output_monophones -l output_log_file output_dict input_dict1 input_dict2 ... input_dictX

---------------------------------------------------------------------------------------------------

In this example, we will run the command like following,

---------------------------------------------------------------------------------------------------

$ HDMan -m  -w lists/wlist -n lists/monophones1 -l dlog dicts/dict1 dicts/beep dicts/names

---------------------------------------------------------------------------------------------------

First, check the input files which load in by HDMan.

• ./lists/wlist
  

The file stores the words for our training setence.In other words, the training sentence is constructed by the words.

這個檔案儲存了我們要進行訓練的字詞。也就是我們的訓練句子是由這些字組成的。

=============================

CALL

DAVE

DIAL

EIGHT

FIVE

FOUR

JULIAN

LAW

LEE

NINE

OH

ONE

PHIL

PHONE

SENT-END

SENT-START

SEVEN

SIX

STEVE

SUE

THREE

TWO

TYLER

WOOD

YOUNG

ZERO

=============================

• ./dicts/beep

The Pronouncing dictionary is built by the professional organization. We can use the information to build one own dictionary.

專業的字典檔案，我們可以利用這個檔案來建造自己的字典。

=============================

# British English Example Pronciation dictionary

#

# Version 1.0

#

# Friday 22 August 1997

#

# Commercial use is prohibited.

# See the file README for more details

#

# Tony Robinson ajr@eng.cam.ac.uk

#

A ah

A ax

A ey

A-BOMB ey b oh m

A-BOMBS ey b oh m z

A-LEVEL ey l eh v l

....omitted.

=============================

• ./dicts/names

The above dictionary may not much enough for our training words.

上面的字典可能不夠足夠。

For example, TYLER t ay l ax, the word doesn't appear in beep dictionary.

Or, JULIAN jh uw l y ax n, the word appear, but the pronouncing doesn't. 

像TYLER並沒有出現在beep裡，JULIAN的另一個發音也沒有在beep裡。

所以針對我們的訓練資料，要檢查字典是否完備，否則日後訓練可能會有辨識錯誤的情況產生。

=============================

DAVE d ey v 

JULIAN jh uw l y ax n 

JULIAN jh uw l ia n 

LAW l ao 

LEE l iy 

PHIL f ih l 

SENT-END [] sil

SENT-START [] sil

STEVE s t iy v 

SUE s uw 

SUE s y uw 

TYLER t ay l ax 

WOOD w uh d 

YOUNG y ah ng

=============================

In order to make sure all neams will have pronunciations, we use "./dicts/names" to save pronunciations of all names.

Actually, there is another one file which will be loaded. It's global.ded.

• global.ded
  

It's the defualt parameter of HDMan, option is -g if you want to load another dictionary file to be global.ded.

=============================

AS sp

RS cmu

MP sil sil sp

=============================

AS commands - AS A B ... - Append silence models A, B, etc to each pronunciation.

$ AS sp

• AS會把sp附加在每個phone sequence後面。

RS commands - RS system - Remove stress marking. Currently the only stress marking system supported is that

used in the dictionaries produced by Carnegie Melon University (system = cmu).

$ RS cmu

• RS會拿掉重音標記，現在只支援拿掉cmu的重音系統。

MP commands - MP X A B ... - Merge any sequence of phones A B . . . and rename as X.

$ MP sil sil sp

• MP會把 sil sp變成sil。

To know more, please refer Section 17.5 in HTKBook.

Following is the commands of HDMan, but a little different one in HTKBook.

---------------------------------------------------------------------------------------------------

$ HDMan -m  -w lists/wlist -n lists/monophones1 -l dlog dicts/dict1 dicts/beep dicts/names

---------------------------------------------------------------------------------------------------

We will generate 3 files and load 3 files by above command (global.ded is default load).

• -m, merge the dicts/beep and dicts/names
• -w, check the words in lists/wlist
• -l, ouput the process log dlog

The 3 input files have been shown.

Here are 3 output files.

• ./lists/monophones1

This is a list of the phones used. 這是有出現在wlist裡的words，有參考到的phones。

其實就是在dlog檔裡面有顯示的phones。

==============================

k

ao

l

sp

d

ey

v

ay

ax

t

f

r

jh

uw

ia

n

y

iy

ow

w

ah

ih

sil

s

eh

th

uh

ng

z

==============================

• ./dicts/dict1
  

Now, we get the dictionary for our training words.

現在我們針對遜鏈字詞產生了新的對應好的字典。

We can see the words in wlist with the phones.

可以看到有出現在./lists/wlist的words，後面都有了自己的phones。

Because of the AS command in global.ded, we can see each words with the sp to be the last phone.

因為AS指令，所以每個word後面都有sp結尾。RS和MP的效果，在這裡都看不出來。

Format is like the pp.26 in HTK book.

WORD [outsym] p1 p2 p3 ...

For more detail, please refer Section 12.7 in HTKBook.

===========================

CALL            k ao l sp

DAVE            d ey v sp

DIAL            d ay ax l sp

EIGHT           ey t sp

FIVE            f ay v sp

FOUR            f ao sp

FOUR            f ao r sp

JULIAN          jh uw l ia n sp

JULIAN          jh uw l y ax n sp

LAW             l ao sp

LEE             l iy sp

NINE            n ay n sp

OH              ow sp

ONE             w ah n sp

PHIL            f ih l sp

PHONE           f ow n sp

SENT-END        sil

SENT-START      sil

SEVEN           s eh v n sp

SIX             s ih k s sp

STEVE           s t iy v sp

SUE             s uw sp

SUE             s y uw sp

THREE           th r iy sp

TWO             t uw sp

TYLER           t ay l ax sp

WOOD            w uh d sp

YOUNG           y ah ng sp

ZERO            z ia r ow sp

===========================

• dlog

The log file includes the statistic of phones used also includes the information of the loaded dictionary file, ./dicts/beep and ./dicst/names. 

If you wish to remove the WARNING message in dlog file, you can manualy create the empty beep.ded and empty  names.ded first.

===========================

WARNING: no script file dict/beep.ded

WARNING: no script file dict/names.ded

Dictionary Usage Statistics

---------------------------

  Dictionary    TotalWords WordsUsed  TotalProns PronsUsed

        beep    237211         23     256476         25

       names        12         12         14         14

       dict1        35         26         39         18

26 words required, 0 missing

New Phone Usage Counts

---------------------

  1. k     :     2

  2. ao    :     4

  3. l     :     8

  4. sp    :    27

  5. d     :     3

  6. ey    :     2

  7. v     :     4

  8. ay    :     4

  9. ax    :     3

 10. t     :     4

 11. f     :     5

 12. r     :     3

 13. jh    :     2

 14. uw    :     5

 15. ia    :     2

 16. n     :     7

 17. y     :     3

 18. iy    :     3

 19. ow    :     3

 20. w     :     2

 21. ah    :     2

 22. ih    :     2

 23. sil   :     2

 24. s     :     6

 25. eh    :     1

 26. th    :     1

 27. uh    :     1

 28. ng    :     1

 29. z     :     1

Dictionary dict/dict1 created

===========================

Practice to add words in wlist file, ex adding HOWARD and CHOU, save as wlist1.

Because HDMan can't find CHOU in dictionaries beep and names.

We just get the HOWARD hh aw ax d sp, and without CHOU.

You also can discover this in log file.

練習自己新增words在wlist，例如新增HOWARD和CHOU，另存為wlist1，

若字典不完備，使用HDMan產生的dict1裡，就只會產生

HOWARD          hh aw ax d sp

卻不會有

CHOU 

Log檔也會顯示

28 words required, 1 missing

若沒發現，後續的訓練語句出現CHOU，就無法辨識成功。

Practice 2, adding Howard and Chou, save as wlist2. After executing HDMan, there are two missing, that caused by HOWARD in beep is not equal to Howard in wlist2.

再練習，新增Howard和Chou，另儲存為wlist2，執行HDMan，log檔會發現出現兩個missing，

28 words required, 2 missing

這是因為Howard 不等於HOWARD，所以要注意。