Documentation – Basic Administration Library

4 Improving text-to-speech

Circumreality relies heavily on text-to-speech. Unfortunately, text-to-speech

is of questionable quality because the AI and simulation needed to

produce quality text-to-speech is a long-long ways away.

However, you can take some steps to improve text to speech:

1. 
   
   Use less text-to-speech - I know this may sound strange,

   but minimizing the amount of synthesized speech helps. One of the reasons

   I added voice-chat to Circumreality was to minimize the amount that text-to-speech

   was used.

   
   
   



2. 
   
   Don't mix text-to-speech with real speech - One sure

   way to make text-to-speech sound lousy is to put it side-by-side

   with real recorded speech. However, if you still wish to use

   recorded speech, MIFL will let you do it.

   
   
   



3. 
   
   Use transplanted prosody - This technology takes a recording

   of your voice, "copies" the prosody (pitch, timing, and volume) from the

   recording and "pastes" it onto the synthesized voice. It significantly

   improves the quality of text-to-speech, but it's a lot of work,

   and is only useful for pre-canned phrases. Automatically generated phrases

   can't be pre-recorded for text-to-speech.

   The speech log is useful for transplanted prosody. See below.
   
   




4. 
   
   Use automatic transplanted prosody - See below.
   
   



5. 
   
   Use quick transplanted prosody - See below.
   
   



6. 
   
   Create a prosody model - Part of the reason why text-to-speech

   sounds so bad is because its prosody (pitch, timing, and volume) are so

   wrong. Prosody is wrong because the text-to-speech system doesn't really

   understand what it's speaking, so it doesn't know what words to emphasize.

   
   
   

You can "Create a prosody model" specific to the text your game world uses.

It will require you to record 500-1000 phrases that are commonly spoken

by the text-to-speech engine, and then learns the prosody for those

phrases, or ones similar.

The speech log is useful for transplanted prosody.

Also, the Easy recording tool might come in handy. See below.



7. 
   
   Make your own TTS voice - Making your own text-to-speech voice

   is a lot of work (about 5000 recorded phrases), but will help for two reasons.

   
   
   







1. In the act of recording your own text-to-speech voice, you end

   up creating a prosody model, as above.
   
   





2. Text-to-speech synthesizes a word, like "lantern", by attaching

   smaller recordings it copied from other words. For example: "lantern"

   may be constructed from the "lan" part of "land" and the word "turn" that

   were recorded in the training sentences. If a recording of the word

   "lantern" exists in the training sentences, then that will be used

   in preference to "lan(d)" + "turn". A recording of the original word

   will sound better than combining two smalelr bits.

   However, "lantern" may not actually have been in the text-to-speech

   training set, or may not have occurred enough to be considered important

   enough to keep. If you create your own voice using the 5000 most-common

   phrases used in your game world, the words that you

   most commonly use will sound better

   as a result.

   
   
   




The speech log is useful for transplanted prosody. See below.

The speech log keeps track of the 5000 most-commonly used

text-to-speech phrases. You can use this list of phrases as (a) a

way to decide what sentences to record for transplanted prosody,

(b) sentences to record for your prosody model, or (c) sentences to

record for your own text-to-speech voice.

To see the list of phrases, type "Show speech log" using

your administrator account. This will display the most common 2500 phrases,

sorted by the most common ones first. (This assumes that your world has

been running long enough to actually speak 5000 different phrases.) It doesn't

display the last 2500 since they have probably only occurred once or twice,

and aren't statistically signifcant.

As a programmer, you should be aware of the following:

• 
  
  oSpeechLog is the object that stores the speech

  log. It is a sub-class of cSaveToDatabase, so it will be

  saved away and reloaded.
  
  




• 
  
  oSpeechLog.SpeechLog() is called whenever SpeakNarrator()

  and SpeakObject() are called. If you bypass these two functions you

  may wish to append call oSpeechLog.SpeechLog() yourself.
  
  




• 
  
  gSpeechLogMax controls how many spentences will be stored

  in the speech log. You can lower or raise this value. 5000 is a

  good value, even if you are creating your own text-to-speech voice, since

  many of the sentences you record for text-to-speech should be

  "phonetically balanced" (look it up on the internet) and not taking

  from your game's phrases.

  
  
  

The chances are, you'll be play-testing your IF title and hear text-to-speech

speak a phrase that ends up sounding lousy because of the prosody.

You would always run M3DWave.exe and record the wave then and there, but

there's an easier solution:

1. When you first run CircumrealityWorldSim.exe, check

   the Enable tools in the client checkbox at the bottom of the

   page. This will enable some author-specific tools in the client that

   are normally hidden from typical players.
   
   





2. In the client, select the "Settings & Options" menu item,

   following by the "Speech settings" dialog. In

   that dialog, type in a Record-wave-to directory where

   you want any recorded waves to be saved, such as "c:\\MyDirectory".
   
   


If you're recording waves for your own text-to-speech voice then

type in the directory where all the TTS recordings are saved. If

you want to the use wave for a prosody model, type in that

directory. Or, if you're using automatic transplanted prosody (see

below), then type in the directory where all those files go.


3. Whenever you hear a text-to-speech phrase that has lousy

   prosody, bring up the transcript window. Next to

   each sentence, you'll see a red link for "(Record)". To

   record a version of that sentence, press (Record).

   The file will automatically be saved in the "Record-wave-to

   directory".
   
   


If you are using the wave file for a text-to-speech voice or a

prosody model, you'll still need to run MNLP.exe

and add the recording to the list of wave files

used for the voice or model. Assuming that the directory where

you saved the file is only meant for the specific text-to-speech voice or

prosody model, you can press the "Add all visible wave files" at the

bottom of the wave-file open dialog box.

Likewise, if you're using the wave file for transplanted prosody,

you'll need to rebuild the automatic transplanted prosody resources.

The MIFL editor makes it easy to record transplanted prosody for strings and

spoken strings in some resources (such as conversation scripts, cut scenes, and

speak-scripts.)

Whenever you modify a string, or other appropriate resource, you'll see

options for "Spell check", "Speak", "Record transplanted prosody", and

"Delete trans. pros.".

• 
  
  "Spell check" makes sure all of the words are in the

  text-to-speech voice's vocabulary. If they aren't, they'll probably be

  mispronounced, and you'll need to add them using 3D Outside the Box's

  "Natural Language Processing" application.

  
  
  



• 
  
  "Speak" has the text-to-speech engine speak the

  phrase without using transplanted prosody. This will help

  identify words not in the voice's vocabulary, as well as detect

  gramatical mistakes.

  
  
  



• 
  
  "Record transplanted prosody" lets you easily record a

  transplanted prosody resource for the string. The resource is named,

  "rTransProsQuick_XXX", where XXX is a randomly generated number.

  Whenever you use Speak(), SpeakScript(), SpeakNarrator(), etc., the

  transplanted prosody that you just recorded will be used.
  
  


This is the easiest way to use transplanted prosody.

• 
  
  "Delete trans. pros." deletes a rTransProsQuick_XXX resource

  for the given text.

  
  
  

If you're trying desperately to improve the quality of a text-to-speech

voice, you should first try to use "Easy recording" (above) to record

wave files for a prosody model.

However, a prosody model may not be able to speak a specific sentence with

the correct and subtle prosody. If that's the case then the next step

is to use "Automatic transplanted prosody".

To use "Automatic transplanted prosody", do the following:

1. Use "Easy recording" (above) to record problem sentences as .wav files.

   Store them all in the same directory.
   
   




2. 
   
   Create a new library in your project where you will store

   the automatic transplanted prosody resources.

   
   
   




3. In the Misc menu, select the Automatic

   transplanted prosdy tool.
   
   




4. 
   
   Fill in all the fields, especially the directory where

   you have saved all the troublesome .wav files.

   
   
   




5. Press Create transplanted prosody resources from .wav files and

   the computer will process for awhile. (If you have just recorded some

   new .wav files, this may take a long time.)

   
   
   




6. Switch to your List of resource, also in the "Misc" menu.

   
   
   

You will see a number of new resources labelled "rTransProsAutoXXXXX", where

XXXXX is replaced by a number. These are the transplanted prosody resources

automatically generated by the .wav files you stored in the directory.


7. When you have a NPC SpeakScript(),

   or the narrator SpeakNarrator(), if

   the text exactly matches the transplanted prosody text, then

   the transplanted prosody resource will be spoken instead of the text.
   
   


You should be able to hear the difference in quality. If you're not 100%

convinced, then have the NPC speak the text string, but with an added space,

so the automatic transplanted prosody resource won't be used.

5 Log-on notices

Showing a notice when players log on.

Log-on notices

If you want to show a notice when players log on, such as, "The mountains

are under rennovation. Please E-mail me if you find any bugs," then

do the following:

1. Override rLogonMainNotice with your own notice.

   
   
   




2. Change gNoticeVersion to a new number so that players

   will be shown the notice. (A notice will only be shown once, and the

   unique ID in gNoticeVersion is used to determine if the notice has

   already been seen.)
   
   

Download 8.87 Mb.

Share with your friends: