There is significant potential clinical benefit to be gained in capturing symptom data from individuals with Parkinson’s Disease (PD). For this purpose, sensor data is often collected. However, labels (ground truth) data is also beneficial, both to train (supervised learning) and to validate outcomes from automated monitoring systems. With the increasing use of voice assistants, this modality has been proposed for labelling. In this study, we examine some design patterns for voice-agent-supported labelling, identify failure modes, and make use of the MDVR-KCL dataset to benchmark a widely used key component, a speech-to-text pipeline. We identify that this component shows rapid increase in several error metrics (WER, CER, WIL) when employed on data from mildly symptomatic participants. We identify some potential mitigating steps and discuss potential future work.