Meta's AI Translations Inference
summary
Pipeline performs autodub and lipsync at Meta scale using the Seamless model (six languages) plus 10+ auxiliary models for language ID, VAD, ASR, noise extraction and alignment, with sentence-splitting and a time‑stretching algorithm to match translated audio length. Production constraints drive model orchestration, streaming interfaces to avoid uncompressed frame transfers, and scheduling to meet low latency. Safety uses watermarking, provenance metadata and runtime toxicity checks, and the system reports ~90% eligibility success with plans to add more languages, music tracks and multi‑speaker dubbing.
AI-written summary · about
chapters (9)
- 0:00AI translations at Meta
- 0:59AI in media processing
- 2:25AI translation pipeline
- 3:48Audio translation workflow
- 8:46Safety in AI translations
- 11:00Lip syncing for viewing
- 14:31Translation quality measurement
- 16:57Production model orchestration
- 17:59User feedback and improvements
related talks
45mTranscend Language Barriers with Custom Machine Learning APIs: um...
33mMaximize lifetime value of media content
16mBeyond Translation — Protecting and Scaling Global Processes with AI Speech Technology
44mScaling Semi-Supervised Production-Grade ASR on 200 Languages
24mApplications of AI in news and media
43mBreaking the language barrier: how do we quickly add multilanguage support in our AI application?
16mThe End of Awkward AI Transcriptions
13mMining Parliamentary Gold - Building Hebrew ASR