The hurdles we overcame
We started our work by dividing our team into separate groups with each one working on a different aspect of the project.
The bluetooth team : we faced some huge obstacles in integrating the new smart watch developed by the design team into the app. We worked to upgrade the modules in the android application. Our final goal was the make the entire process seam less so that when there was an impulsive sound in the environment, the band would immediately trigger the app. The user could then use the app to distinguish the type of sound.
The algorithm and model training team : we began our work by reading multiple different research papers on different kinds of Convolutional and Recurrent Neural Network models for sounds classification. We narrowed down on the final model keeping in mind the accuracy as well as optimising the final computation that would be required. However our problems had just begun, there are multiple speech classifiers which already exist however very few sound classifiers. This meant that we had to manually scour the web for sound files for a large training database. In the end we had found more than thousands of training samples. Once we finished this, we reached the most difficult step of the project, the algorithm. We spent many nights debugging and modifying many different codes which could help generate the final graph which we would integrate in the application.
The android app development team: As we were all new to the world of android development we spent our first few weeks learning to work on Android studio and creating basic applications from scratch. We then moved onto updating the previous application which had been developed for SmartBand. We fixed crash issues as well as updated the older APIs which had been used. The major issue we faced was in designing a new part of the application which could support Tensorflow so it could support the new sound recognition model provided by the algorithm team. Google simultaneously began deleting many older Tensorflow library sources as they planned to introduce a newer version of Tensorflow lite. This made our work more difficult as we had to figure out many roundabout techniques to integrate the previous libraries.
The bluetooth team : we faced some huge obstacles in integrating the new smart watch developed by the design team into the app. We worked to upgrade the modules in the android application. Our final goal was the make the entire process seam less so that when there was an impulsive sound in the environment, the band would immediately trigger the app. The user could then use the app to distinguish the type of sound.
The algorithm and model training team : we began our work by reading multiple different research papers on different kinds of Convolutional and Recurrent Neural Network models for sounds classification. We narrowed down on the final model keeping in mind the accuracy as well as optimising the final computation that would be required. However our problems had just begun, there are multiple speech classifiers which already exist however very few sound classifiers. This meant that we had to manually scour the web for sound files for a large training database. In the end we had found more than thousands of training samples. Once we finished this, we reached the most difficult step of the project, the algorithm. We spent many nights debugging and modifying many different codes which could help generate the final graph which we would integrate in the application.
The android app development team: As we were all new to the world of android development we spent our first few weeks learning to work on Android studio and creating basic applications from scratch. We then moved onto updating the previous application which had been developed for SmartBand. We fixed crash issues as well as updated the older APIs which had been used. The major issue we faced was in designing a new part of the application which could support Tensorflow so it could support the new sound recognition model provided by the algorithm team. Google simultaneously began deleting many older Tensorflow library sources as they planned to introduce a newer version of Tensorflow lite. This made our work more difficult as we had to figure out many roundabout techniques to integrate the previous libraries.
Comments
Post a Comment