Hi,
In my current model: Emergency vehices alarm - Dashboard - Edge Impulse
I am trying to set features in the way, that model will evaluate ony certain frequency band .
I understood that this cold be done by setting 2 parameters of “Log Mel-filterbank energy features”: 1) Low frequency 2) High frequency
However the fields, where I am supposed to type a value are greyed out and initial value 0 can not be changed. Am I doing something wrong, or this can not be changed for Syntiant Audio parameters setting?
Is there other way, how to set the frequency band for ML model?
I did not mentioned that I need to deploy the model into Syntiant TinyML board. My understanding is that Syntiant Audio processing block is the only one that is compatible with this HW. I am not sure if Spectral Analysis Processing Block can be used in this case.
You must use the Syntiant Processing Blocks. See this post. I assume when @jbuckEI refers to DSP blocks he is refering to what the EdgeImpulse Studio calls Processing Blocks.
Maybe jbuckEI can tell us why the fields are grayed out.
Do I understand correctly that you want the output of the Processing Block (yes @MMarcial that is what I meant) to output features only in the band of interest? If so, I can see how not being able to change the frequency settings is a limitation. I will have to look into why some values are fixed, though I suspect it has to do will how the NDP101 is designed.
There are a number of limitations for the NDP101, please read this example documentation carefully.
I’ll update more when I know, and get some more info into the docs.
The Arduino Nicla Voice and the Syntiant TimyML boards both use a SoC that integrates a MEMs mic with a DSP that converts the signal to PDM. The PDM signal directly feeds the Syntiant AI Sensor (NDP101 or NDP120).
So there is no way to add a circuit to filter the analog signal.
It might be possible to put in a CPU between the Mic and Syntiant AI Sensor to create a filter on the PDM data.
Yes, that’s how I understood it. However, unless there is another way to implement the filter without any HW modifications, I will have to rely on the ML model, which hopefully will learn to omit unwanted low-frequency noise.