audio decompression in realtime?

[Zamuel_a]

Is it possible with some audio decompression in realtime for audio samples? Had been nice to have that in a game, but it has to be rather fast then and compress the sample to some amount. If its only to 80-90% its not worth it.
Is there else some way to rather fast decompress a audiosample, not in realtime, so it just saves space on the disc, but can be decompressed rather fast anyway.

[Zorro 2]

Only Cybernetics do that ! (hum... I think)

There are a little production only for Atari STE : Flexi Glass

The official website : http://cyber.savina.net/

[bod/STAX]

It's possible.

I think Gunstick/ULM did it in his parallax distorter screen in the Dark Side of the Spoon demo using delta-compression.

[Zamuel_a]

Ok, I did some testings with delta compression. From what I know delta compression is made by you take one sample and you calculate the difference from the previous so if you have 3 samples in a row with the same value you get a delta value of 0 and if you have 2,4,6,8 you get a delta value of 2,2,2. I tried that and could only compress a sample to about 10% perhaps. I had to resample to 7 bits to, so I could have one bit telling me if I had more than one sample with the same delta value, or else I had to (in worst case) spend 2 bytes for every byte of sample. If I just did a compare between the samples and grouped samples of same value, run length encoding I think it's called. I could actually compress more than with the delta technic, atleast with the test sample I had, so I don't know really how to do it to gain anything much.

[earx]

zamuel: you're doing lossless compression. lossy crunches much harder with the risk of making it sound like shit. it all depends on the predictor and amount of quantisation. google for ADPCM.

predictor = guess for the next sample. often this is just taking the current sample. but there are loads of smart choices for specific cases.

quantisation = this basically is the precision of the amplitude. you can quantise to 4 bits to cover a range of 16, but also a range of 32 (taking steps of two)

there are smart ways like Sony's soundchips in the PSX (VAG compression, a bit inefficient on 68000) or a cheaper one of which i forgot the name.
i tried a VAG derivative on the Falcon's DSP and it rocks.

[Nyh]

Is it possible with some audio decompression in realtime for audio samples? Had been nice to have that in a game, but it has to be rather fast then and compress the sample to some amount. If its only to 80-90% its not worth it.
Is there else some way to rather fast decompress a audiosample, not in realtime, so it just saves space on the disc, but can be decompressed rather fast anyway.

The ARJ M7 depacker can depack samples at a rate of about 64 kB/s on a normal ST. In most cases you can improve the compression ratio by delta encoding the sample before packing and add the values before playing.

The trick with ARJ is good for lossless compression. To get better compression ratios you can do a lossy compression. Use something like mu-law encoding. If you make it table driven it can be pretty fast. When you encode at nibble level you can use the sequence: -128, -64, -32, -16, -8, -4, -2, -1, 0, 1, 2, 4, 8, 16, 32, 64. With two 256 byte tables you can decode the upper and lower nibble without any shifting. If the quality is not good enough you can add some bits. The decoder becomes a bit more tricky and time consuming as you have to decode a bit stream.

Hans Wessels

[earx]

MrNi: LZ coding is as good as useless on more-or-less noisy data. Huffman or another entropy coder do help. I tried a pure Huffman codec and zip and the differences were (logically) negligible.

[Nyh]

MrNi: LZ coding is as good as useless on more-or-less noisy data. Huffman or another entropy coder do help. I tried a pure Huffman codec and zip and the differences were (logically) negligible.

That is correct. The compression ratio on samples of an LZ coder is not impressive. The test file I used left 80% of the original. That is why I advised to use a delta encoder before using an ARJ. Even then you will not get impressive results. But on loss less compression it is about as good as it gets with the standard tools available on the ST.

Hans Wessels

[unseenmenace]

I don't know exactly what it is or if its any good to you but theres something labelled as "Sample Packer" on Medway Boys menu 74

[Gunstick]

It's possible.

I think Gunstick/ULM did it in his parallax distorter screen in the Dark Side of the Spoon demo using delta-compression.

oh? someone mentioned ULM?

*raising from the digital grave*

Yes, I think there has been no screen in ULM history featuring digisound which was not compressed. Packing twice the sound in the same space is just too appealing.
And you DON'T HEAR the difference! At least not on an YM.

Most amazing is that it is possible to play compressed sound faster than uncompressed.
For more infos on dominotable check this thread viewtopic.php?f=1&t=4357&p=32070&#p32070 also read some posts above that about why a table on 2, 4, 8, ...128 is better than any other way to do it.

Georges
PS: that entire thread is worth a read for the technophiles.

[atari_mark]

Boris of POV did something like this on our Invisible Man demo and the Ivor Biggun demo. The source code may be on the POV CD but I can't check at the moment because I am at work. Boris' routine was a bit difference one that only worked on some samples, we could play a sample and load/depack on the fly in the background.

Mark

[Zamuel_a]

I tried to do some testings with compression like Gunstick did on the parallax distorter screen with the information here http://www.atari-forum.com/viewtopic.php?f=1&t=4357&p=32070&#p32070
But I can't get it to work. I made a test sequence of data and then compress it to 4 bits and then back with the table 1,2,4,8... technic but I can't get it to work. It looks ok at the beginning of the sequence but then the errors get higher until it gets useless.

First I created two tables:

Code: Select all

BYTE delta_conversion_table[256];
char delta_table[16] = {0,1,2,4,8,16,32,64,127,-64,-32,-16,-8,-4,-2,-1};

void createDeltaTable()
{
int i,j=-128;

   for (i=0; i<255; i++,j++)
   {
      if (j==0) delta_conversion_table[i] = 0;
      if (j==1) delta_conversion_table[i] = 1;
      if (j>=2 && j<=3) delta_conversion_table[i] = 2;
      if (j>=4 && j<=6) delta_conversion_table[i] = 3;
      if (j>=7 && j<=12) delta_conversion_table[i] = 4;
      if (j>=13 && j<=24) delta_conversion_table[i] = 5;
      if (j>=25 && j<=48) delta_conversion_table[i] = 6;
      if (j>=49 && j<=96) delta_conversion_table[i] = 7;
      if (j>=97 && j<=127) delta_conversion_table[i] = 8;
      if (j>=-128 && j<=-97) delta_conversion_table[i] = 8;
      if (j>=-96 && j<=-49) delta_conversion_table[i] = 9;
      if (j>=-48 && j<=-25) delta_conversion_table[i] = 10;
      if (j>=-24 && j<=-13) delta_conversion_table[i] = 11;
      if (j>=-12 && j<=-7) delta_conversion_table[i] = 12;
      if (j>=-6 && j<=-4) delta_conversion_table[i] = 13;
      if (j>=-3 && j<=-2) delta_conversion_table[i] = 14;
      if (j==-1) delta_conversion_table[i] = 15;
   }
}



After that I calculate the delta values like this:

Code: Select all

void calcDelta()
{
int   i;
BYTE   previous = 0;
BYTE   delta;
BYTE   c;

   for (i=0; i<buffer_length; i++)
   {
      c = buffer[i];
      delta =c - previous;
      delta_buffer[i] = delta_conversion_table[(BYTE)(delta+128)];
      previous = c;
   }
}


"buffer" is a unsigned char buffer with the data to compress

And to decompress it I do this:

Code: Select all

void decompress()
{
int   i;
BYTE   index;
char   value;
char   old_value=0;
BYTE   sample;

   for (i=0; i<buffer_length; i++)
   {
      index = delta_buffer[i];
      value = delta_table[index];
      sample = old_value+value;
      old_value = sample;
      buffer[i] = sample;
   }
}


I can't see any errors. I have tried most char/unsigned char combination if it's a overflow problem with negativ and positiv numbers, but this is what gives something close to a result. I might have missunderstod the whole thing of how to do it?

I made a testtable:
1 10 14 150 4 216 59 24 200 18

and then I decoded it I got:
1 9 13 140 11 235 106 74 10 74

So it looks ok at first, but then halfway in it gets all wrong.

[Nyh]

So it looks ok at first, but then halfway in it gets all wrong.

I didn't do a thourough analisis of your code but I think you go wrong here in the void calcDelta() function.

I would say:

Code: Select all

void calcDelta()
{
  int   i;
  BYTE   previous = 0;
  BYTE   delta;
  BYTE   c;

     for (i=0; i<buffer_length; i++)
     {
        c = buffer[i];
        delta =c - previous;
        delta_buffer[i] = delta_conversion_table[(BYTE)(delta+128)];
        previous += delta_buffer[i];
     }
}

With this code you are working with the calculated value to determine the next delta, as you do in your decompressor. With your code the errors keep adding up until there is no correlation left between input and output.

Hans Wessels

[Zamuel_a]

Thanks! It actually works now!
I had to replace that last line to:
previous += delta_table[delta_buffer[i]];
since the delta_buffer just keeps index numbers and not the real values. But then I changed it to that I actually got something out that sounded more or less like the sample I converted. There are some "clicks" here and there in the sample, but I guess that could be

[Zamuel_a]

because of some error I still got in the code.
It sounds much better than then I tried to resample to 6 bits, but now it's only 4!

[Gunstick]

I did not read all your code. What you have to do is, calculate delta, then get the 4 bit value nearest this delta, then add that value to the current sample to get the next theoretical sample (you "previous").
This is the same as letting running the decompressor at the same time with the compressor so that the compressor creates the waveform as close as possible to the original.
And don't forget that 254+4=2 which makes loud samples sound great.