Mimigaraoboeru
It's just a simple content-based recommendation system. I built it for a flashing idea that if I can find Japanese songs with lots of new words, good melodies and good lyrics. See 1.png in detail.
Required platform, libs and datasets
Python 3.6.* Mecab 0.996 mecab-python ffmpeg tensorflow keras
FMA Dataset
Usage
Launch word_process.py to find all new words in words.csv (JLPT-N1 words list). Launch craw_lyrics to craw song lyrics on NetEase Cloud Music's server; put favourite song lyrics in ./lyrics/0_fav/ and name them like <number_of_times_heard——recently>.txt (e.g. 560108_70_MEMORIA.txt) Launch filtrate_lyrics to filtrate non-Japanese characters in all the lyric files. Launch theme_bias.py to get the theme bias(the likelihood that the lyrics of the song will fit you) of each song and dump it to theme_bias.pkl . Launch music_download.py to download songs that have high theme bias(too large if download all of them). * the variable t_ represents the number it will download. Launch wave_render.py to do STFT(Short-Time Fourier Transmission) on downloaded .mp3 files. Launch music_bias.py to predict the feature of each .mp3 file and get the music bias by calculating their cosine similarities. Launch theme_bias.py to get the final order due to theme bias, music bias and lyric importance. Launch download_netease.py to download the top rated songs of fianl_order.pkl .
** You may change some codes to finish all of those steps. Create an issue if you need help. I may do engineering optimization if I have enough time.