Supporting data for the paper: R. Hennequin, J. Royo Letelier, M. Moussallam "Audio Based Disambiguation Of Music Genre Tags", ISMIR 2018

The repo gathers csv files that contains embeddings vectors for genre tags and similarity matrices between music genre tags for the 3 experiments done in the paper:

• the taxonomy_learning folder contains embeddings $f_w$, $f_c$, $f_m$ and $f_{\text{dist}}$ and related similarity matrices for the taxonomy learning experiment described in section 3.1 of the paper. Tags are the 250 most popular Discogs genre/style tags.

• the deduplication folder contains embeddings $f_w$, $f_c$ and $f_m$ and related similarity matrices for the deduplication experiment described in section 3.2 of the paper. Tags are the 250 most popular Discogs genre/style tags that were duplicated as explained in the paper leading to 500 duplicated tags.

• the mumu_discogs_translation folder contains embeddings $f_w$, $f_c$, $f_m$ and $f_{\text{dist}}$ and related similarity matrices for the translation experiment described in section 4 of the paper. Tags are the 250 most popular Discogs genre/style tags and the 211 most popular MuMu genre tags .

Some large csv files were compressed as zip files to limit size. Moreover, the size of $f_c$ (and for $f_{\text{dist}}$ in the taxonomy learning experiment) was limiter by sampling 20000 samples in the test dataset to avoid generating very big files.

Embeddings can be easily loaded in python using pandas:

import pandas as pd
embedding_df = pd.read_csv("taxonomy_learning/fw_tag_embedding.csv")

# embedding vector f_w for genre tag "acid jazz"
print(embedding_df[u"acid jazz"])

Very similarly, similarity matrices can be loaded this way:

import pandas as pd
similarity_df = pd.read_csv("taxonomy_learning/fw_tag_similarity.csv", index_col=0)

# similarity based on f_w between tags "heavy metal" and "thrash"
print(similarity_df.loc["heavy metal","thrash"])

The repository also contains the datasets that were used to train the CNN and to compute the embeddings as 3 csv files (train.csv, validation.csv and test.csv) for each experiment, containing deezer song Ids and labels. Metadata and audio previews for each song IDs can easily be retrieved using the Deezer API.