The voice acting in beta4 is terrible
This model has some inexplicable lines that completely disregard the MH3 cues, making the voice acting a jumbled mess. The character even reads out all the cues, which is ridiculous. It's a failed model.
While he's already acknowledged that beta 4's audio is not good, the behavior you describe is a common failure of even the Minimax base model. Be careful not to include audio if there is no speech or clearly defined vocalizations, and make sure actions and sound are enough to fill the clip length.
It seems I made a mistake. Changing the reference node in ComfyUI's built-in MiniMax H3 fixed the issue, I used the MiniMax H3 Director plugin node, thinking it would make output easier, but it caused various noises and dubbing malfunctions, and the character's movements were reduced in size and accuracy. This plugin is really unreliable, everyone should be careful, I now only use it to standardize the prompts, it's useless for anything else.