1

I'm trying to build an intrusion detection device by training a model using footsteps sounds.
I'm trying to do it using TensorFlow.
The problem is that this example does not work at all: I've tried to create a custom word called "footsteps", created a directory in the right place with wavs files in it (with the same format of the example's wavs) and, even if the script finishes running, it does not produce an output file (it prints a path where the output file should be, but the directory does not exists) nor a confusion matrix (it gets printed as 'None') and it says that the validation accuracy is 0%.

I've tried to change approach, starting from the micro_speech example of the TensorFlow lite library.
Even here, I've used this file to produce a new feature vector. It works fine, but it fails to recognize the footsteps 99% of the times (obviously, since I've "trained" the "model" with just one wav file).

Is there a way to train that example with more files, or a simpler way to do what I need that I somehow missed?
Thanks in advance.

Riccardo
  • 47
  • 6
  • Not sure speech recognition is the right way to detect non voice sounds. As far as I know (but I night be wrong) speech recognition is designed for noises, that can be generated with a 'model' of a vocal chord and with the related speech spectrum. Just for fun try to train it with the same code but with a spoken word first. So you know whether it's due to your code or due to the fact, that the noise of foot steps doesn't work with speach recognition. – gelonida Nov 30 '19 at 14:57
  • Thank you, I will try that, the problem is that speech recognition examples were the only examples I was able to found so I decided to give it a shot. – Riccardo Nov 30 '19 at 15:21

0 Answers0