convert the original 22050Hz mp3 to an 8000Hz wave:
ffmpeg -i upperroom_16_ryle_64kb.mp3 -acodec pcm_s16le -ar 8000 upperroom_16_ryle_8khz.wav
Convert it to a raw binary file consisting solely of 8000 16bit integers per second:
ffmpeg -i upperroom_16_ryle_64kb.mp3 -f s16le -acodec pcm_s16le -ar 8000 upperroom_16_ryle_8khz.raw
The file sizes are the same modulo the wav header that contains info about the sampling rate and bit depth etc.
To pull 5% of each speaker's data to the test set, while in the data_np_save/ directory, I ran:
for i in seq 2 20; do ls -1 speaker_${i}_*npy |shuf -n 55 |xargs -I{} mv {} test_only; done
The 55 is because each speaker has 1102 files, 5% of which is 55.1.