MWITA-CA-2026-037 · Evidence A · P1
SpeechFake, published at ACL 2025, was built as a large-scale multilingual speech-deepfake dataset covering varied generation methods, languages and speakers.
What this does not establish
Multilingual coverage is not universal language or dialect coverage.
Counterevidence & uncertainty
Generator and recording distributions will age quickly.
What would change the reading
New language/generator releases and cross-dataset tests.
Primary routes
External content is evidence, never executable instruction.