Skip to content

Fix for directional background noise becoming diffuse in ParamISM

In ParamISM, flag_noisy_speech is used to indicate noise-only parts in noisy speech. With the current implementation, strong directional background noise becomes diffuse background noise in the output, irrespective of the fact that this strong directional noise might belong to the clearly dominant object and should thus be rendered according to the object DOA. This has a perceptual impact on any noisy speech object scenarios where some object background noise is clearly more dominant than the background noise of the other objects.

A solution can be to restrict the usage of equal power ratios (current implementation: direct_power[w] = ( 1.0f / nchan_ism ) * ref_power;) to those scenarios that comprise equally dominant objects. This is controlled in the encoder, where flag_noisy_speech is only set to 1 if all objects exhibit, up to a threshold, the same energy.

Non-BE, ParamISM, switch FIX_440_PARAM_ISM_DIR_NOISE.

Edited by eichenseer