Audio-to-Speech produces incorrect output when pose conditioning is disabled (enable_skeleton_cross_attn = False)

[Salaheldinyoussry]

Thank you very much for your excellent work on this project and for sharing it openly.

When running the model in audio-to-speech mode without pose conditioning, setting
enable_skeleton_cross_attn = False causes the model to produce unexpected / incorrect output.

The frame size is 1:1

RD_Radio12_000_frame_out_video.mp4