DSP result error

Hi friends,

I am here trying to learn this edge impulse platform, so I was trying to make a sound classifier model and uploaded my data sets. But when I carried forwarded with recommended “syntiant” the DSP result section prints an error “Error: Target size (512) must be at least input size (1687)”. Can someone help me what am I missing or what wrong I am doing?

Thanks in advance

Hi @ayushkr316,

This block is only for Syntiant hardware target only that requires a specific window size. I’ll check with the team to remove it as “recommended”.
I’d suggest using MFCC (for speech recognition) or MFE/Spectrogram if you are trying to classify non-humans speech.

Aurelien

Hi Aurelien,

Yeah, Thanks for the response. I have very limited knowledge in this field and hence I chose the recommended block. Although I selected MFCC in next try and it worked well.

Ayush

1 Like

Glad to hear it!
If you want to learn more about processing blocks, you can read our docs on Spectrogram/MFE and MFCC. For generic keyword spotting the default MFCC parameters should work well.

Aurelien

Thanks Aurelien, Sure I will read them for enhancing my skills. I was just curious that if it is a good idea to use MFCC and MFE both processing block together for better results? What do you suggest?

Ayush

MFCC is the same as MFE, with an additional processing step that is chosen based on the properties of human speech. Thus, you should just use one or the other based on Aurelien’s advice: MFCC: human speech. MFE: non speech sounds

Hi AlexE,

Thanks for the information. I was using non human speech + noisy background with MFCC, but now it seems MFE is a better choice for my application.

And thanks, we’ve fixed the recommended flag too.