S/M misclassifications on deep voices

Basic info

• Commit SHA: main
• Platform: All

Bug description

In the recent ISM1 test we have observed degradations on some input items containing deep male voices. A closer analysis revealed that this is due to occasional but systematic mis-classifications occuring in the second stage of the S/M classifier, in the function ivas_smc_mode_selection().

Ways to reproduce

./IVAS_cod -ism 1 NULL 13200 48 input_file.wav bit

Unfortunately, the input file cannot be shared but it should be possible to use other appropriate item. The mis-classifications can be observed either with the sp_aud_decision1 parameter or the core parameter.