Keyword Spotting Hardware

Hello! this is Ahmad Naeem, I am very new here so please pardon me if my quesiton has already been answered.

I am working on a project which requires to identify and perform certain action when a particular keyword is spotted. Total number of keywords that I need to identify are 190 which I know is a very huge number.

I am at very early stage of the project trying to discover proper hardware for it. What should I do?

1- Should I go to the heterogenous computing where some keywords will be detected by one controller and other by the other controller due to the memory constraints?

2- Or should I go to the SBC and develop my application on platform like Raspberry Pi?

I want to use Edge Impulse for my project, help in this regard is greatly appreciated.

Thanks in advance
Ahmad

@Ahmad01

What type of project (industrial, thesis, …)?

The number of keywords is large, so that it would be a challenge. Of course, it depends on the accuracy you like to target.

I suggest exploring transfer learning and using the Edge Impulse Python SDK with TensorFlow and Keras for the design of the model. Once you have a model, you can profile and explore which is the best target device for your application.

Regards,
Joeri

1 Like

Thanks @Joeri for your reply, we are creating a commercial device so we can say it is a industrial project and as per requirements we need accuracy of 98 percent outside of the cabin and 94 percent inside the cabin (any information about cabin is classified I can’t share).
I understood your point, now after profiling if I land on the target device that is not supported by Edge Impulse what would be roadmap then?

Regards

@Ahmad01
Once you have a model, you can generate it to a generic C++ library using ei.deploy that can be used as starting point for your firmware application integration on an own board.

I suggest getting in contact with Louis Moreau @louis. He can bring you on the right track.

Regards

@Joeri
It is really helpful for me, thanks again.
I will try to get in touch with @louis

Regards