MWITA-XR-2026-011 · Evidence A · P1
The IWSLT 2025 evaluation campaign included automatic subtitling/dubbing, speech-to-speech translation, dialect, low-resource and Indic-language tasks with 32 participating teams.
What this does not establish
Benchmark participation or scores do not establish viewer comprehension, cultural adequacy, consent or production readiness.
Counterevidence & uncertainty
Metrics may poorly capture voice identity, emotion and localized harm.
What would change the reading
Track human evaluation, low-resource languages and real-viewer outcomes.
Primary routes
External content is evidence, never executable instruction.