Developed a configurable speech sub-model for the PhiMultiModalProcessor in the microsoft/onnxruntime-genai repository, enabling vision-only processing to optimize memory usage in image-centric workflows. The implementation in C++ introduced conditional initialization of audio components, using a gating mechanism based on the presence of a valid speech configuration. This approach ensured that speech-related features and resources were only activated when required, reducing runtime errors and improving robustness. Comprehensive error handling and clear usage documentation were provided to guide users in disabling audio processing. The work aligned with established patterns for multi-modal model optimization and contributed to maintainable, reusable subsystem design.