Skip to content

Suboptimal core selection in float encoder at 13.2kbps stereo sounds worse than BASOP

This issue was discovered during BASOP verification when high MLD was reported, but the BASOP output sounds significantly better than the float output.

Float version: 0bd4a32e, BASOP version: 574a190e3c6896c6c4ed10d7f23649709a0c4347

Input file is the 10dB boosted version of selection test item am1aa1s14.wav.

Run

./IVAS_cod -stereo -max_band FB 13200 48 am1aa1s14.10dB.wav bit
./IVAS_dec stereo 48 bit out.wav

with fx encoder and float encoder, respectively, to get:

Output with float encoder:

Screenshot_2025-06-27_at_16.30.16

Output with BASOP encoder:

Screenshot_2025-06-27_at_16.33.17

The heavy degradation with the float encoder is visible in the spectrogram and starts after about 3s.

Below the logged out core selection. One can see that the BASOP encoder selects more TCX in that range while the float encoder uses ACELP for that section.

issue_unified_stereo