On July 17, at the World Artificial Intelligence Conference (WAIC), Baidu made its inaugural presentation of the digital human video podcast solution, One-Mirror. This innovative solution has achieved a breakthrough in digital human micro-expression technology, enabling natural interruptions and harnessing the expressive power of professional film and television audiovisual language. It effectively tackles challenges such as the high costs associated with real-person interviews, the rigid expressions of traditional digital humans, the disjointed flow of multi-person dialogues, and the distortions that occur when stitching together different modalities. In doing so, it paves the way for the industrial-scale production of in-depth interview content.

From a technological standpoint, the solution sets up three robust layers of defense: the ERNIE large model facilitates meticulous planning of micro-expressions, transcending the constraints of single emotion labels; Baidu’s self-developed digital human specialized model refines expression control precision to within 1 second, ensuring seamless synchronization of eye contact, body language, and speech; and through extensive training with professional audiovisual samples, it achieves a profound multimodal unity, significantly elevating the generation capabilities of digital humans.

On the product front, leveraging three AI intelligent agents for screenwriting, directing, and editing, along with a repertoire of over 80 skills, the solution can autonomously generate speculative interview scripts. It supports solo lectures and dual-person opinion dialogues across a diverse range of fields, including finance, law, technology, and humanities. Real-world test data reveals that this solution slashes production costs by 74%, amplifies content expressiveness by 785%, extends the average playback duration by 29%, and outperforms mainstream models on the market in terms of character naturalness.