Skip to content

Very different coding of an attack at 48 kbps for ISM-1

Basic info

  • Encoder (float): 311956f3 (same results with d705aad41)
  • Encoder (fixed): 45f78445
  • Decoder (float in both cases): 311956f3 (same results with d705aad41)

Bug description

The beginning of a speech burst is completely different between float and fixed point . Actually the fixed point looks more like the original, while the floating point is going straight to clipping. Issue observed at 5.5s. In the figure below, from top to bottom:

  • Original
  • Floating point synthesis used for fixed point development (sha 311956f3, float-float)
  • Fixed point encoder, floating point decoder
  • Latest floating point synthesis (sha d705aad41, float-float)

image

Ways to reproduce

G4S1.wav.0.csv

G4S1

IVAS_cod_flt -ISM 1 G4S1.wav.0.csv 48000 48 G4S1.wav bit.flt

IVAS_cod_fx -ISM 1 G4S1.wav.0.csv 48000 48 G4S1.wav bit.fx

IVAS_dec_flt EXT 48 bit.flt syn.flt.wav

IVAS_dec_flt EXT 48 bit.fx syn.fx.wav

To me it looks more like a float problem, tagging @malenovsky and @vaclav for awareness.