Intel Music Coder

From MultimediaWiki
Revision as of 14:21, 30 October 2006 by Roberto (talk | contribs) (no longer incomplete)
Jump to navigation Jump to search
The printable version is no longer supported and may have rendering errors. Please update your browser bookmarks and please use the default browser print function instead.

Perceptual audio codec developed by Intel. It features a 256-point MDCT transform and constant bitrate (all known samples are mono and have 512-bit frames). These coefficients are grouped into 32 bands, the first bands hold 3-4 coefficients and the last bands contain 12 coefficients per band. Bit allocation for each coefficient is done adaptively for each frame. The window for MDCT overlap is a simple sine window normalized with sqrt(2).

Frame Structure

All bits are packed into 16-bit little-endian words, MSB first.

 9 bits - should always be 0x21
 3 bits - frame type
 1 bit  - flag (used in bit allocation)
 <band coefficients>
 <band flags>
 <coefficient flags>
 <coefficients>

Band Coefficients

Each band should have one coefficient. If the frame type has bit 0x4 set then the first band coefficient should be read explicitly otherwise it should be read using one of 16 Huffman trees depending on frame type and band number (frame type specifies one of 4 sets which contain four static Huffman trees and an array pointing to one of those trees for each band number). If the returned Huffman code is 17 then read 4 additional bits and add their value to the result.

Band Flags

Each band except bands with coefficient == 16 should have its own flag.

Coefficient Flags

Coefficient flags are stored either as bits or may be coded as pairs:

  0 -> 1, 1
 11 -> 0, 1
100 -> 1, 0
101 -> 0, 0

Coefficients

Coefficients are stored in 1-6 bits with length calculated depending on bits left and bits per band.