The most significant risks that could jeopardize the success of this project would be our inability to tune the motor controllers correctly so that they can properly give the right instructions to the motors themselves. We’re currently working on this portion of the project, and once that’s done, we should mostly be done with the project, apart from incorporating the audio component to handle pivoting to the designated actor after the previous actor’s dialogue is completed. Once that is done, we’ll be ready for the TechSpark demo along with the final demo, and we’ll also be able to meet on Thursday with Dr. Dueck and the School of Music for our demo with them for the “Ah Perdona” piece. Apart from that, we will of course also be working on the final report and the final video so that we can submit all of them in time when they are due.
On the audio side, the main tests that we performed were calculating the WER for the ASR system, specifically for the trigger phrases rather than the whole dialogue. We realized that it didn’t make sense to track the whole dialogue by each speaker, rather it would improve processing and accuracy if we only focused on the few words before the next actor’s part began. What we found was that the system was very accurate, with a 7% WER (10% is high-quality) on these trigger phrases, and that the trigger was almost always correctly initiated. However, the fact that it wasn’t perfect concerned us slightly, which is why we included a manual override button in the UI that allowed the director to control the cueing themselves in case the trigger didn’t work. On the tracking side, the main tests we performed were calculating the difference between the ground truth and the predictions by the UI, and we discovered that the accuracy was even better than we expected, allowing us to stick with our current design. On the overall system side, we tested the reliability of the entire system, and we confirmed that the nodes and other components all ran for at least 4 hours, which was what we were hoping for.