According to a report by the webmaster's home on December 22, Meta has recently released a series of AI translation models, which achieve real-time voice conversion latency of no more than 2 seconds, support multiple language translations, and have the ability to imitate characteristics such as tone, speech speed, and emotion. This family of models, called Seamless Communication, includes SeamlessExpressive, SeamlessStreaming, SeamlessM4 T v2, and Seamless, the first three of which have been open-sourced on GitHub. To ensure translation accuracy and avoid abuse, Meta employs toxicity mitigation technology that filters out "toxic content" before training and automatically detects and adjusts the generated toxic words during translation generation, while watermarking the audio to trace the source. To prevent the risk of abuse, Meta has also added a watermark to the audio, which allows you to accurately trace the source of the audio and combat various attack vectors by embedding an imperceptible signal in the audio.