I’m using the edge-impulse-uploader CLI tool to upload a collection of 1-second audio clips (.wav, 32-bit, 16 kHz) to a project, and every file is appearing as a blank audio file in the project’s “Collected Data.” I get no sound when I play the file in the browser, and when I download the file, it’s just a 1 second blank piece of audio. I did verify that there were sounds in the file prior to uploading.
I’m creating a dataset based on the Google Speech Commands dataset (http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz), if that offers a good starting point to test this.
Any help would be appreciated!
Hmm… Me and @dansitu are using the same dataset for some work. Could you email me one of the files that does not properly upload (jan@edgeimpulse.com)? I’ll take a look!
@ShawnHymel, the files in that dataset were encoded using the WAVE_FORMAT_IEEE_FLOAT format, instead of the PCM format we were expecting (WAVE files look so simple at first glance!). I’ve pushed v1.4.12 of the CLI that fixes this. Files in this format will be downsampled to 16-bit (losing a bit of depth), but it shouldn’t matter too much for performance. On-device you probably want 16-bit mono sound anyway to run inferencing fast.
Awesome, thank you! On-device, I’ve been using 16-bit samples with 16 kHz, but it seems that the Google Speech Commands Dataset use 32-bit 16 kHz samples, so I truncate to 16 bits during training.
1 Like