MWITA-CA-2026-038 · Evidence A · P1
SpeechFake’s experiments test how generation method, language diversity and speaker variation affect detector performance rather than assuming one aggregate score transfers across conditions.
What this does not establish
A headline accuracy without cross-condition breakdown is insufficient field evidence.
Counterevidence & uncertainty
Real telephony and social-platform transformations may differ.
What would change the reading
Codec, noise and adversarial field tests.
Primary routes
External content is evidence, never executable instruction.