Generation Provenance Before Behavior Attribution: Auditing Synthetic Speech Research Objects

TL;DR

This paper proposes a generation-provenance substrate to audit synthetic speech research objects, enabling the attribution of model behavior to synthetic training data. The substrate binds source specification, generated content, and other relevant information. It contributes a compact provenance contract and an audit protocol.

Why It Matters

Understanding the provenance of synthetic speech data is crucial for attributing model behavior and ensuring the reliability of AI systems. This research has significant implications for the development of trustworthy AI models, particularly in high-stakes applications like healthcare. It highlights the need for transparent and auditable data generation processes.

Key Findings

• generation-provenance substrate is necessary for behavior attribution
• proposed substrate binds multiple relevant information types
• audit protocol is demonstrated in a private Japanese care-handoff pipeline