Skip to content

[Regression] MR1026 introduces regression in SBA decoding

Bug description

The MR !1026 (merged) has introduced a small regression in the SBA decoding. It's the same issue as reported in !871 (comment 59822) for the difference BASOP - Float:

After !1026 (merged):

Bildschirmfoto 2025-01-15 um 16.47.40.png

Before:

Bildschirmfoto 2025-01-15 um 16.52.52.png

Commandline:

  • IVAS_dec binaural 48 HOA3.dm4aa1s14_10dB.enc_48kHz-SWB-384000-flt.192 out.wav

The bitstreams + decoded files are on box.com, ...\Box_EXTERNAL_IVAS_BASOP_VERIFICATION\issues\issue-1101

In the course of !871 (merged) @tyagiri , @jiaquan.huo and @tritthart have already done some analysis, but we honestly don't fully understand it - the new code is supposed to be better since it avoid the saturation, but the difference still gets bigger. Our best explanation is that there is another issue somewhere else, which causes this problem.