ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
     V9t9:  TI Emulator! v6.0 Documentation      (c) 1995 Edward Swartz
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
  ÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÍÍÍÍÍÍÍÍÍÍÍ SPEECH.TXT ÍÍÍÍÍÍÍÍÍÍÍÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ

       This file details V9t9's speech synthesis capabilities.


  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
                        ³ THE EMULATED SPEECH ROM ³
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ

       First of all, the emulated speech ROM file "EMUSPCH.BIN" has no
  words in it and is practically useless.  It is only meant to allow
  programs with their own speech to function.  (These programs will not
  use speech unless they find some sort of speech ROM.)


  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
                         ³ V9t9 SPEECH SYNTHESIS ³
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ

       In earlier versions of V9t9, a call to the speech synthesizer would
  either flash a word up on the screen, or "speak" it through a Sound
  Blaster -- by reading its digitized waveform of disk.  This digitized
  file took up 1.5 megs, and in hindsight was a very silly thing to do.
  That file is no longer supported and is obselete.

       LPC speech -- linear predictive coding -- was made exactly for the
  purpose of storing speech in very low amounts of memory.  The word
  "HELLO" takes up about 48 bytes of memory, when encoded into LPC
  equations.  Digitized at 8000 Hz, it takes about 8000 bytes.  So...


  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
                 ³ SO, LPC IS INFINITELY BETTER, RIGHT? ³
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ

       Theoretically.  I've implemented LPC synthesis into the emulator,
  doing away with that 1.5 meg file entirely.  However, since I only got
  it working recently, it's not perfect yet.

       There is a whole table of values in the speech synthesizer that
  holds the key to my correctly implementing LPC speech, but I know of no
  way to access it, or how to properly guess at the values contained
  therein.  So, the speech you hear is off a bit, and there's nothing I
  can do about it but guess some more.


  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ
                   ³ IT SOUNDS HORRIBLE ON MY COMPUTER. ³
  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ

       Same with me.  It may be because your computer is too slow.  After
  trying for weeks to figure out the reason for the bad results on my
  system, I ran V9t9 on fast computers and realized that the speech
  actually sounds pretty good.

       Try running V9t9 off a clean boot.  EMS memory managers really slow
  down speech.

       This bad-speech-on-slow-computers can be considered a bug.  I took
  a half-assed approach of setting the interrupt rate to 8000 Hz when
  generating the speech, though a DMA-based approach would be much faster.
  I'll fix it in later versions.


  ÍÍÍÍÍÍÍÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ÍÍÍÍÍÍÍÍÍ
         ³ WHAT IF I KNOW HOW THE COEFFICIENTS ARE REPRESENTED? ³
  ÍÍÍÍÍÍÍÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙÍÍÍÍÍÍÍÍÍ

       Well, then *stop reading* and e-mail me!  Hee hee.  To avoid
  confusion, I know exactly how many bits the equations use, and in what
  order they are, and all that junk.  That's why you can hear something
  remotely intelligible.

       What I do not know is how the 6/5/4-bit LPC coefficent indices are
  mapped into the corresponding 10-bit fixed-point coefficients (range
  negative 1.0 to positive 1.0).  I've tried linear, logarithmic, and
  exponential transformations but none of these quite cut it.

       The inaccuracies may be due to my incorrect implementation of the
  inter-frame parameter interpolation, the guessed-at "chirp" sequence, or
  the guessed-at pitch translation table.  (I'll fix the interpolation in
  the next version when I'm not afraid to completely break the code for a
  little while.)  However, I doubt these account for the remarkably bad
  results I get.

       I think TI used a standard implementation of LPC-10, which was
  standardized by the Federal Government, some year between 1976 and 1979.
  Of course, I can't find information on it.

       The contents of the speech synthesizer's "Parameter Look-UP ROM" is
  exactly what I need.  If you happen to know the values, or how to access
  them from the speech synthesizer, or you worked for TI, or you actually
  created the speech ROMs (hee hee), feel free to inform me about this.


  ÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ