Deezer Research

Next week, #Deezer research team members will be attending #RecSys2024 in #Bari with many exciting presentations and talks planned

Bruno Sguerra, research scientist in our team will be giving a talk at #MURS https://lnkd.in/gvqh_BFB a workshop co-located with the ACM RecSys 2024 Conference. #Recsys2024

This week at INTERSPEECH 2024, Yuexuan Kong presented her research paper on the Dataset Strada. Joint work with Viet-Anh TRAN and Romain H. at #Deezer If you are interested in discovering the work and the detailed Singer Traits Dataset that comes with it, the paper is accessible: https://lnkd.in/eKzQXt_X

Greetings, This year, we have two papers accepted for the ACM RecSys 2024 Conference https://lnkd.in/gwiqZs2e The first one, "Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation," explores the intersection of cognitive science and transformer models. In this paper, we introduce PISA, a model that builds user representations based on the dynamics of memory access provided by the ACT-R cognitive architecture. We leverage both short- and long-term representations of users' tastes to improve next-basket recommendations. You can find the paper here: https://lnkd.in/eNBBT8nS. The second paper focuses on #reproducibility. In "Do Recommender Systems Promote Local Music? A Reproducibility Study Using Music Streaming Data," we evaluate whether classic recommendation models are biased toward international music or if they favor local artists. We reproduced a previous study, where researchers showed that different recommender systems exhibit algorithmic biases, either promoting or disadvantaging local content. However, we used data collected from Deezer, and our results differ significantly from the original study, highlighting how the "quality" of the data influences the findings. More information can be found here: https://lnkd.in/evbKPv3F. Both will be presented by team members in Bari in october. See you there!

Paper news! Two submissions by #deezer team members have been accepted at ISMIR Conference: From Real to Cloned Singer Identification by Dorian Desblancs, Gabriel Meseguer Brocal, Romain H. and Manuel M.: https://lnkd.in/eu62eUu2 where we investigate ways to identify artist's cloned voices in mixtures. STONE : Self-supervised Tonality Estimator by Yuexuan Kong, Vincent Lostanlen Gabriel Meseguer Brocal Stella Wong, Mathieu Lagrange and Romain H. https://lnkd.in/ed9zmZ2t STONE is the first self-supervised tonality estimator using the ChromaNet architecture. ChromaNet produces a 12-dimensional key signature profile, trained by identifying pitch transpositions between unlabeled musical excerpts, correlating with tonal key signatures. STONE is then extended to output a 24-logit KSP, incorporating supervision to distinguish major and minor keys. We show that the semi-supervised version achieves the same accuracy as the supervised version with less supervision and surpasses it with equal supervision.

Our Head of Research Romain H. will be at Wallifornia to talk about #AImusic and the latests work from the team on the topic. https://lnkd.in/eH8745fi

Deezer Research member Elena E. is co-organizing #NLP4MusA 2024, the 3rd Workshop on #NLP for #Music and Audio, co-located with the ISMIR Conference in San Francisco later this year. If you're doing #research at the intersection of music, audio, and language processing, including (but not limited to) language modelling for music understanding, retrieval or generation, we encourage you to submit your work! Submissions are open until 5th July. 🔗 Workshop website: https://lnkd.in/enj3hdA9 📅 Submission deadline: 5 July 2024 📅 Workshop date: 15 November 2024 🌉 Workshop location: San Francisco (ISMIR 2024)

We are excited to announce that our paper "STraDa: A Singer Traits Dataset", work by Yuexuan Kong, supervised by Viet-Anh TRAN and Romain H., was accepted at INTERSPEECH 2024 . This work contains two main contributions: 1. The release of Singer Traits Dataset (STraDa), which contains two subsets: automatic-strada and annotated-strada. Automatic-strada contains 25k downloadable tracks across numerous genres and languages of more than 5k unique lead singers, as well as cross-validated lead singer metadata and track metadata. Annotated-strada contains 1200 segments of 200 tracks balanced across two genders (all singers considered in this work are cis-gender), four age groups and five languages. 2. To show its use for model training and bias analysis, we benchmarked singer sex classification and conducted bias analysis. We believe Strada offers a unique opportunity for model training and bias analysis for singing-voice-related tasks. We warmly invite future researchers to use and enrich Strada for various tasks. preprint of the paper is available on arxiv: https://lnkd.in/ewjv_gVn #mir #publication #dataset

Last week Gaspard Michel and Elena E. visited Avignon University to present their latest work on NLP. Thanks Yanis Labrak for organizing it and everyone there for the warm welcome. https://lnkd.in/eyPVk6vc

Last week, Dorian Desblancs and Gabriel Meseguer Brocal were at #ICASSP2024 to present our latest work on Self-supervized methods for music tagging https://lnkd.in/eYDM-89S It's been picked by Jordi Pons i Puig among his prefered contributions of the conference: https://lnkd.in/ediw_Pzs