Meta’s latest large language model, Llama 2, is trained on two trillion tokens. Yet, with 90% English training data, it “may not be suitable for use in other languages,” Meta warns.
Meta AI introduces Voicebox, a multilingual, multitasking generative AI model for speech, trained on 50,000 hours of audio and 20 times faster than previous models.
In just three months, universities and tech companies released over two dozen papers on speech-to-speech translation (S2ST), highlighting continuing trends, new directions in research.
Researchers from Meta suggest training-data curation as a means to reduce added toxicity in machine translation. Toxicity categories include profanity, insults, hate speech, bullying, etc.
Interpreters pushback on RSI, Meta’s MT “breakthrough,” growth in media localization with ZOO and Keywords, Zoom’s translation feature, and Disney experiments with synthesized voices.
The new model, praised by Meta founder and CEO Mark Zuckerberg, is already being used to improve more than 25 billion translations daily on Facebook, Instagram, and other apps.
Meta eliminates text generation in speech to speech translation; shares code and research with the public.
Slator Weekly: Join over 15,800 subscribers and get the latest language industry intelligence every Friday
Tool Box: Join 10,000 subscribers for your monthly linguist technology update.
Your information will not be shared with third parties. No Spam.
This will close in 0 seconds
