Voice Activity Detection (VAD) is a root component in Human-Robot Interaction (HRI), especially for use cases such as a self-learning personalized conversational robot partner designed to support elderly users with high acceptance. While state-of-the-art, lightweight deep-learning–based VAD models achieve high precision, they often struggle with low recall in environments with significant background noise or music. In contrast, traditional lightweight rule-based VAD methods tend to yield higher recall but at the expense of precision. These limitations can negatively affect user experience, particularly among elderly individuals, by causing frustration from missed spoken inputs and reducing overall usability and acceptance of the conversational robot partners. This study investigates noise-suppressing preprocessing techniques to enhance both the recall and precision of existing VAD systems. Experimental results demonstrate that effective noise suppression prior to VAD processing substantially improves voice detection accuracy in noisy settings, ultimately promoting better interaction quality in elderly-centric robotic applications. Moreover, optimal sample rate, frame duration, thresholds and voice activity modes were identified for the robot Double3—the conversational robot partner platform for seniors in a care home, co-creatively developed by reflecting with the nursing staff. An open-source dataset and a dataset collected and annotated in-house with the Double3 robot were evaluated for robustness in benchmarks.