Cloud-based AI – an experiment

During Milestone’s partner symposium, MIPS, in Hanoi, this year, Hans Jørgen and his team decided to test how well cloud-based AI really works. All the experiment required was a camera with a microphone, a Google Cloud account, Internet access and some software. In one of the experiments, the team transmitted the audio from the camera to the Google AI service 32 times per second and received a transcript of the data. Thereby they could make automatic subtitles, which was stored in a metadata stream for future searching.

The team concluded, among other things, that the location of the microphone impacts the quality of the transcription. Moreover, the test showed that the transcription does not start until a full sentence is registered.

“One of the major challenges for cloud-based AI is the network – high continuous load bringing enormous amounts of data to the cloud triggers a not insignificant bandwidth cost.  A further challenge is that high latency may course door not to open fast enough if access is based upon face recognition,” said Hans Jørgen Skovgaard.

Read the full blog in Danish here.