TooLAME 02l - 2 March 2003

• Major psychoacoustic model overhauls
  • For more detail, see psycho models
  • Psychoacoustic Model 3
    • Psycho3 is a reimplementation of the psychoacoustic model 1 from the ISO standard. Pretty much totally rewritten from the ground up, following the nomenclature of the ISO standard.
    • Uses arrays for keeping tone/noise labels rather than the really hard to grok pointer-stuff from the dist10 code.
    • Uses Painter + Spanias' Formula for the ATH rather than the tables from the standard.
    • Uses LAME's freq-to-bark conversion to construct the critical bands.
    • Future: Needs a proper geometric mean for weighting the noise. Needs one more function added to the decimation routines which would eliminate maskers within 0.5 dB of each other.
    • Eliminated all the ISO/dist10 tables. Everything is from equations or built from scratch.
  • Psychoacoustic Model 4
    • A reimplementation of psychoacoustic model 2.
    • Eliminated all the ISO/dist10 tables. Everything (ath, bark, critical bands) is from equations or built from scratch
    • FUTURE: For psycho model 2 and 4 there's some really bad "warbling" and "Davros" type noise. Depending on the loudness of your sound sample, this can get really annoying. I don't know where it's coming from.
  • Psychoacoustic Model 0
    • This model uses the ATH and the scalefactors for each subband to build an approximate SMR for each subband at nearly zero cost.
    • Based upon an idea mentioned in "Low power mpeg/audio encoders using simplified psychoacoustic model and fast bit allocation" by Hyen-O Oh et al.
    • For the amount of effort that this psycho model puts in, the results are pretty good.
    • Future: Add some parameters to the equation to allow it to be tweaked on the fly.
  • Psychoacoustic Model -1
    • This is the old "fast" psychoacoustic model ("-f").
    • All it does is copy over a static set of pre-calculated SMR values
    • Sounds OK for most stuff.
• New bitstream encoding routines (encode_new.c)
  • All the old tables.c/alloc_table stuff is now superfluous.
  • All tables are now at the top of encode_new.c
  • The tables have quite a bit of indirection to get to the value that you need (but really not any more indirection than the old alloc_table stuff). Probably need to add some more docs to say what's going on.
  • These new routines are the default, but you can remove the "NEWENCODE" definition and use the old ones (just in case).
  • The new routines will become the default in the next release.
• More speed
  • All the trig stuff for psychomodel 4 is now done with tables instead of calculating exact values.
  • The exact trig values aren't really used directly in the encoder. They're sort of averaged over a couple of iterations and used as a predictor of uncertainty. So being off a few thousandths won't really affect anything.