In the wave of enterprise digital transformation, meeting efficiency has always been a pain point. Traditional meeting minutes rely on manual work, which is not only inefficient but also results in information retention rates of less than 30%. The AI voice solution jointly launched by GeekDance and Dting aims to completely solve this problem through a software-hardware integrated design.
Hardware Breakdown: The Geek's Persistence
To achieve ultimate sound pickup, we customized a 6+1 circular microphone array. But that's not enough; we embedded a dedicated DSP chip on the edge, performing millisecond-level beamforming and sound source localization locally before uploading the audio signal to the cloud.
Edge-Cloud Collaborative Architecture
Traditional pure cloud-based recognition suffers from latency. We adopted a hybrid architecture of "edge preprocessing + cloud large model analysis." The edge handles VAD (Voice Activity Detection) and noise reduction, while the cloud's trillion-parameter Transformer model is responsible for complex semantic understanding and meeting minutes generation.
Automated Workflow
After the meeting ends, the system not only generates a verbatim transcript but also automatically extracts "to-do items (Todo)" and pushes them to Feishu/DingTalk. This is a typical application of AI Agent in office scenarios.
Technology is not the goal — exponential business growth is.
