Koemei is a spin-out of the Idiap Research Institute created in 2010 pioneering an automated rich transcription platform operating on heterogeneous data like lectures, group meetings and video-conference. The company has been selected in the Swiss Top100 startups in 2011. Since its creation, Koemei has gained substantial market in this business; however a number of non-Swiss partnerships, activities and contracts have been slowed down by the strong CHF. In such a situation, it is a priority for Koemei to reduce from one side its delivery time and computing costs and on the other side to increase the quality of its services in order to gain market over competitors.
The goal of this project (DIMHA) is to assist Koemei development and foster its international competitiveness by integrating an advanced audio segmentation and speaker diarization software into its platform through the transfer of IM2 second phase technology. Speaker Diarization is an essential and time consuming step in the rich transcription process that can negatively affect costs and quality. In last four years, in framework of the second phase of IM2, Idiap developed a novel speaker diarization toolkit based on a new non-parametric framework. The system can achieve better then state-of-the-art results running faster then real-time. Furthermore it can easily handle multiple feature streams with a very reduced computational cost, an essential capability to increase the quality of the rich transcription process. This technology perfectly suits the company's need in this moment. DIMHA is organized in two tracks 1. Diarization Software Optimization and Integration and 2. Multiple-Stream Diarization for Koemei Heterogeneous Data sets. The main challenge for DIMHA will be moving the last IM2 research outcomes into solutions for Koemei's ongoing and starting contracts. DIMHA is expected to have immediate effects on the company and especially produce cost savings and delivery time reduction in order to protect the company from the strong CHF in its international activities. Furthermore, the
quality improvements brought by the multi-stream modeling will permit Koemei to accept contracts on very large and heterogeneous datasets, widening his business opportunities and gaining market over its competitors.