[ Bonk lossy/lossless audio coder ]

Version 0.5

Written by Paul Harrison

Bonk is a high quality audio compression program. It can operate in
either lossless or lossy mode. In lossless mode, the exact original
WAV file can be recovered from the compressed file. In lossy mode,
some information is discarded in the compressed file, yielding a much
higher compression ratio.  The information discarded is perceptually
unimportant, and the result should be a *perceptually* lossless encoding.
Bonk can compress some types of sounds more than others, so the actual
bit-rate achieved varies.

In lossy mode, it can compress some types of sound to as low as 95 kbps
(a compression ration of 14:1) while maintaining perceptually lossless
CD quality stereo.  In general, the compression ratio achieved by Bonk
is slightly higher than that achieved using the more common MP3 format,
for equivalent sound quality.  In particular it copes with transients
(eg clapping, drum beats) better. Performance on purely tonal sound is
roughly equivalent to MP3.

In lossless mode the compression ratio is typically around 2:1.

Bonk differs from most existing lossy coders in that it does not
use a Fourier transform to analyse the audio. Instead it uses a
linear predictor (as used in speech compression) with a very high tap
count. This approach has several advantages. Firstly, the power spectrum
of a linear predictor is very similar to human perceptual masking, so we
get perceptual masking effects without any extra explicit modelling of
masking effects.  Secondly, it can better reproduce certain sounds such
as clapping which contain transients. Thirdly, it is easily implemented
using fixed point arithmetic, allowing both lossy and lossless modes
using the same algorithm.

Decoding BONK files is computationally expensive, but should be possible
in real-time on computers that aren't too ancient.

Bonk is free software distributed under the General Public License, see the
file COPYING for details.

[ Compiling and installing ]

Bonk currently compiles for Linux, FreeBSD, NetBSD and OpenBSD.  If you
are using NetBSD or OpenBSD, edit the Makefile before compiling (see
the comments in the Makefile for instructions).

To compile and install type

  make
  make install

[ Usage ]

Here are some examples of how to use Bonk:

 * To compress file.wav into file.bonk using the default lossy mode:

       bonk encode file.wav

   The degree of quantization (lossiness) can be adjusted using the -q
   option. This specifies the amount of perceptual loss: Bonk will choose
   an appropriate bitrate based this and the exact nature of the sound,
   higher for more complex sounds, lower for simpler sounds. 

   - high quality

       bonk encode -q 0.5 file.wav

   - default
   
       bonk encode -q 1.0 file.wav

   - low quality

       bonk encode -q 1.5 file.wav

   For speech, first convert to a single channel and an appropriate sampling
   rate (say, 11025 Hz). This can be done with the commonly available "sox"
   program. Then encode using a fairly small predictor size thus:

   - speech, about 20 kbps

       sox file.wav -r 11025 -c 1 new-file.wav
       bonk encode -s 10 -d 1 -q 2.5 new-file.wav

   Note: specialized speech compressors (such as RPE-LTP) can achieve several 
   times better compression.

 * To compress file.wav into file.bonk in lossless mode:

       bonk encode -l file.wav

 * To play a BONK file:

       bonk play file.bonk

 * To decode file.bonk back into file.wav:

       bonk decode file.bonk

Run bonk with no arguments for further usage instructions.

[ Contacting the author ]

My email address is pfh@csse.monash.edu.au. I am interested to know
what you think of the BONK format and program. Are there any features
missing which should be added? Are there particular types of music it
doesn't cope with? Do you want to give me lots of money so I can devote
more time to work on it?

[ Change log ]

0.1 - Initial release

0.2 - Problem with predictor which caused a warbling effect for very pure
        tones fixed (by limiting the impulse response of the predictor)
      Energy used to determine quantization calculated directly from errors 
        now, rather than estimated
      Neatened up-sampling interpolation

0.3 - Altered up-sampler again (it was cutting off high frequencies)
      Reduced default quantization level
      Extra options:
	Predictor size (for better speech compression)
        Artist, Title (for comment header)
	Output file

0.4 - Endian safe playback (ie, now works on PowerPCs)
      More elegant (and faster) encoding algorithm
      Slightly faster decoding

0.5 - Support for *BSD (courtesy of a patch by Christian Weisgerber) 
      Now automatically increases bitrate for tonal sounds

