Photo by Marvin Meyer on Unsplash
Project Overview: Scaling a Multi-Site Large-Scale Collection
At Meta, I led a multi-month wearable speech data collection project, coordinating 1,000 participant sessions across four sites and owning the operational changes that significantly lifted data quality mid-flight. The core challenge was transforming a deceptively simple request—having participants read text aloud while wearing early-stage hardware—into a high-throughput, rigorous research operation spanning equipment reliability, staffing, protocol standardization, and corpus quality across four and a half months and three escalating phases.
My Role: Lead Research Assistant & Acting Project Manager
As the only person present for the full run of the study from the first session to the last, I acted as the central anchor and project manager for the entire lifecycle of the project. My responsibilities included:
- Personally running approximately 200 data collection sessions across all three phases.
- Onboarding, coordinating, and overseeing a rotating team of 12 Research Assistants plus an external vendor team across four sites.
- Authoring and continuously iterating the day-to-day collection protocol and documentation.
- Partnering with engineering to troubleshoot data infrastructure and streamline collection workflows.
Methodology & Execution: End-to-End Operational Ownership
To transform a high-friction operation into a scalable pipeline, I employed a rigorous, hands-on management approach:
- Direct Session Execution & Continuity: Covered every active collection day while maintaining intimate, ground-level visibility into participant friction points.
- Staffing Infrastructure & Tracking: Built and maintained a multi-site staffing model, tracking per-session attribution in a shared session tracker to catch and resolve coverage gaps proactively.
- Reading-Material Curation: Audited source texts for readability, replacing dense 19th-century prose that caused participant stumbles with a cleaner, modern public-domain corpus to reduce retakes and frustration.
- Infrastructure & Workflow Troubleshooting: Partnered with engineering to diagnose silent upload failures caused by sleep mode, drove the migration to dedicated lab computers, and restructured the scheduling cadence to trade unsustainable raw daily throughput for a sustainable team rotation.
- Protocol Iteration: Continuously updated documentation to govern session flow, equipment setup, consent handling, and hardware handoffs.
Key Findings: Lessons in Scaling Research Ops
This massive undertaking yielded critical insights into what actually drives large-scale hardware collections:
- The first 100 hours were measurably worse than the last 100: Launch pressure front-loaded throughput before workflows were stable; quality was ultimately an infrastructure problem, not a participant problem.
- Reading material is a study variable, not a detail: Text difficulty directly drove participant frustration and session length; optimizing the corpus measurably improved data quality.
- Throughput has a ceiling set by people, not pipelines: The original schedule produced steady burnout; schedules must account for human behavior and fatigue, not just raw metrics.
- Dedicated hardware beats software workarounds: Moving to dedicated lab machines removed an entire failure category rather than just mitigating it.
Outcome & Strategic Impact: Reusable Standards & A 1,000-Hour Corpus
My leadership and operational interventions had a direct, lasting impact on the team’s research infrastructure:
- Higher-Fidelity Data: Mid-flight interventions—specifically the introduction of dedicated lab hardware, the corpus swap, and schedule stabilization—ensured that the back half of the dataset met rigorous production standards.
- Milestone Achievement: Successfully hit the 1,000-hour target, closing out the study with a complete, thoroughly documented dataset.
- Reusable Templates: The staffing and coverage model I built became the template for scheduling subsequent collections, and the protocol I authored is now the official onboarding document for new RAs working on this class of study.
