Continuous Sound Identification using Nicla Voice

[drmpf]

I am using nicla Voice for rain detection with a model build by EdgeImpulse
There are two classes Rain and znoise i.e. not rain
I would like to classifier output to show Rain / znoise
but with continual rain I only get a couple of Rain outputs and then nothing.
phbackoff is 5 for both classes.

This topic is similar

Continuous sound classification

Hi, I am trying to find a solution to problem described on Edge Impulse forum as well: Basically I need to repeatedly classify a sound which is continuous and uninterrupted. When I follow the examples for a keyword recognition the TinyML board provides classification typically on the beginning of such sound and then classification stops - apparently waiting for some change in the sound intensity. I understood that this behaviour might be changed by setting the json posterior parameters (phbac…

but no answer given there.

Questions
1)How do I force continual match outputs?

1)Is there a low level function call to the NDP120 that will return the current clasification result, i.e. classes and their % match in an array/structure

3)What do the posterior parameters
timeout 0
timeout_action “stay”
timeout_action_arg 0
do?
Would changing these make any difference.?

I have already reduced the znoise phth to 30% to better pickup when it is not raining.

[agiri]

Hi @drmpf ,

Thank you for reaching out to us.
We will get back to you soon with all the relevant information.
Regards.

[agiri]

Hello @drmpf

Responding to the queries,

1)How do I force continual match outputs?

Please clarify what is your size of rain detection pattern that you trained with.
Ideally the EdgeImpulse examples as well as the audio examples listed like Nicla Voice, should do continuous inference. Please share some samples of audio files which is not doing continuous match.

2)Is there a low level function call to the NDP120 that will return the current clasification result, i.e. classes and their % match in an array/structure

For Arudino nicla voice, NDP has provided API - NDPClass:poll() which invokes the registered callbacks for the match and you can get the match classification label string results

3)What do the posterior parameters
timeout 0
timeout_action “stay”
timeout_action_arg 0
do?
Would changing these make any difference.?

For the classification problem you are indicating of rain/no-rain, this is not applicable. Please leave it to default