Human communication combines speech with gestures that refer to objects and locations in the surrounding environment. Generative video models offer a promising way to synthesize diverse examples of such behaviour, but visually plausible motion does not necessarily convey the intended spatial reference or align with the relevant words. This project investigates how to evaluate and improve spatially grounded, speech-aligned human video generation. The project builds on MM-Conv, a multimodal conversational dataset in which two-party referential conversations were recorded using VR headsets within a physics simulator, containing speech, motion capture, gaze and annotations of referring expressions. Because the conversations take place in simulated scenes, the intended target, scene geometry and rendering camera are known exactly, enabling controlled experiments. Given a starting image, speech audio and a gesture instruction, video-generation models will synthesize a person referring to a visible object through pointing. We will develop an evaluation pipeline that measures pointing direction relative to the intended target and gesture timing relative to speech. Body and hand motion reconstructed from generated videos will be registered to the known scene. To establish measurement reliability, the reconstruction pipeline will first be evaluated on rendered recordings with known motion, supplemented by controlled spatial and temporal perturbations. Evaluation will account for reconstruction uncertainty, scene changes and visual quality. Experiments will compare pretrained audio-conditioned video models and investigate targeted adaptation to improve spatial grounding while preserving natural movement and speech alignment. Generalization will be assessed across held-out recordings and environments. Following validation, the project will explore synthetic-data expansion across additional scenes and human appearances, and assess its usefulness for referential understanding. The expected outcomes are a reproducible evaluation framework, evidence about the capabilities and limitations of video-generation models for grounded communication, and methods for producing useful synthetic multimodal data. The work supports the development of embodied agents that understand and generate human communication in shared physical spaces. Specifically, the project will address the following research questions: (A) To what extent do audio-conditioned video-generation models produce gestures that accurately indicate an intended object and align with the corresponding spoken reference? (B) Can adaptation using motion recordings rendered in known 3D scenes improve spatial accuracy while preserving natural movement and speech alignment, and do these improvements generalize to unseen environments? (C) Can spatially validated synthetic videos improve referential understanding beyond training on the original recordings or conventional rendered data alone?