Repetitive Sampling Decoder (RSD) is a wickedly fast MP3 decoding library for Python/ML, made with libavcodec (FFMpeg). In the time that it took you to read that sentence, RSD could have easily decoded an hour of audio.
RSD's multithreaded architecture allows the next batches to be CPU decoded in the background while the GPU is busy doing other things. FFMpeg's seek implementation allows for efficient random access of clips from longer audio files.
If you are on debian/ubuntu, you should be able to run
apt install ffmpeg libavutil-dev libavcodec-dev libavformat-dev libavfilter-dev
On arch, this should work:
pacman -S ffmpeg
If you are on other linux or osx, these packages should be available but may have different names.
If you are on windows, may god have mercy on your soul.
pip install numpy
Now you should be able to run make install
Take a look at pysrc/user.py for a minimum viable usage.
Note: RSD is still under active development, so you'll probably see this error periodically even for correctly normalized files.
RSD expects audio files to be uniform in format, sample rate and channel count for maximum decode rate. Use normalize-inplace.sh.