Custom Wake Words Recognition about diy-alexa HOT 2 CLOSED

Apologies, I can see what the issue is.

All the samples need to be exactly 1 second long.

The function def is_valid_file(file_name): checks that there are 16000 samples and rejects files that are not the correct length.

This length limit is due to the way the model works - it treats the spectrogram like a 2D image, so can't deal with longer or shorter samples.

There's a couple of options:

Make sure all your files are exactly 1 second in length
Or:
Modify the code so that it truncates the audio file to 1 second and only rejects files that are too short. You would need to make sure you have at least 1 second of data in the file.

You can modify the def is_correct_length(audio, expected_length): to accept files that are >=16000 samples.

And modify the def process_file(file_path): function to something like:

def process_file(file_path):
    # load the audio file
    audio_tensor = tfio.audio.AudioIOTensor(file_path)[0:16000]

I think that will work - I've not tested it though.

from diy-alexa.

Brilliant ! it's now capturing perfectly ! : )

I trimmed all sound to exact 1 seconds , no more or less.

from diy-alexa.