MIR (Music Information Retrieval) Corpus suggestions

Question

I have been working on a project on MIR.

I have run my code on various datasets like GTZAN, DEAM, EmoMusic

Now I want to train my classifier in a different fashion, i.e. Classification on the basis of Emotion invoked by the song. I tried to use last.fm for this cause but it was not much useful.

Can some suggest me a dataset, where songs are classified on the basis of emotions (Not on the basis of genres pls)?

I know that there is Arousal Valence model, but after going through various Survey papers I have found that the classification using that model is not the best, is there any other mode? or a Dataset someone can recommend.

Thank you.

Answer 1

You are on the right path,so you can either use a dataset providing the valence and arousal levels as meta data or the emotions. The first type is more frequent and easier to find but if you are interested in labeled emotions you can use the circumplex model developed by James Russell also known as Russell's model to convert the valence/arousal levels to emotion labels. (graph source: Perceptually Valid Facial Expressions for Character-Based Applications, https://www.researchgate.net/publication/220061100_Perceptually_Valid_Facial_Expressions_for_Character-Based_Applications)

Here are some of those datasets:

• 4Q audio emotion dataset (Russell's model) (2018)
• The DEAP dataset
• DEAM dataset - The MediaEval Database for Emotional Analysis of Music

As for the datasets with emotion labels:

• The emotify dataset
• Music and emotion dataset (Primary Musical Cues) should be what you are interested in.

More can be found under: https://ismir.net/resources/datasets/