Skip to content

Precison lost during re-scaling of the weight speech memory vector

Basic info

  • Float reference:
    • Encoder (float):
    • Decoder (float):
  • Fixed point:
    • Encoder (fixed): a9c6f8d8
    • Decoder (fixed):

Bug description

The vector old_wsp_fx is unnecssarily rescaled leading to a lost of precision that can affect the computation of some parameters (open loop pitch, voicing,...)

When removing this unnecessary rescaling, the mld score for the ltv48_stereo encoded at 16.4 kbps goes from 4.31 to 2.66 .

Ways to reproduce

Box folder: ...\Box_EXTERNAL_IVAS_BASOP_VERIFICATION\issues\issue-xxxx

IVAS_cod_fx -stereo 16400 48 ltv48_stereo.wav bit.fx.ref
IVAS_cod_fx_no_rescaling -stereo 16400 48 ltv48_stereo.wav bit.fx.new
IVAS_cod_flt -stereo 16400 48 ltv48_stereo.wav bit.flt
IVAS_dec_flt MONO 48 bit.flt syn.flt.wav
IVAS_dec_fx MONO 48 bit.fx.ref syn.fx.ref.wav
IVAS_dec_fx MONO 48 bit.fx.new syn.fx.new.wav
mld syn.flt.wav syn.fx.ref.wav -> 4.33
mld syn.flt.wav syn.fx.new.wav -> 2.66