#### problems
- metadata keys shouldnt be fullpaths but relative ones (imagine downloading this dataset);
- key-pair in metadata doesnt guarantee existence of audiofile;
- instead of writing ffmpeg commands to file maybe better to run them in go;
- change metadata format from json to csv for lower memory consumption;
- add config file; move constance to config file;
- ffmpeg call should be unique;
- ffmpeg commands are slow, split to clips inside of this module;
- wav take a lot of space; no need for conversion, pytorch supports opus (convert to opus);