@electronictechgreg Does it move to noise = 1.0 when you don’t say anything? In that case it could be that your dataset is unbalanced perhaps (if you can give me the project ID I’ll take a look) and the model is a bit underfitting. NN might see some peaks then, but not all the way to 1 when detecting the keyword.
I balanced it out last night as per your video. Project ID 13444.
Thank you for taking a look at my project.
The noise level does go up to .99 when I say nothing. Also, the keywords “Lights On” do not get classified as well as the “Hey Yoda” keywords. Sometimes I do get a correct classification at > .7
Maybe its my code? The testing goes well on the website.
In that case would be very useful to check the audio buffer. If you stitch 4 250ms. windows together does it sound correct? Maybe there’s some clipping / distorted sections there.
The board does have a 3W DAC+PA Audio output. I thought about outputting it to a speaker as you suggested.
I did set my DMA audio buffer to 4000 since my sample rate is 16000.
Also, I set
inference.buffers[inference.buf_select][inference.buf_count++] = (int16_t)(i2s_buf[2 * i]);.
If any of that appears wrong let me know.
Thanks again for your help. I will try the speaker suggestion as I think that might give me a clue.