US 12,242,960 B2
Music release disambiguation using multi-modal neural networks
Aneesh Vartakavi, Emeryville, CA (US); and Konstantinos Antonio Dimitriou, Berkeley, CA (US)
Assigned to Gracenote, Inc., Emeryville, CA (US)
Filed by Gracenote, Inc., Emeryville, CA (US)
Filed on Mar. 16, 2021, as Appl. No. 17/203,631.
Claims priority of application No. 20210100137 (GR), filed on Mar. 8, 2021.
Prior Publication US 2022/0284284 A1, Sep. 8, 2022
Int. Cl. G06F 16/60 (2019.01); G06F 16/65 (2019.01); G06N 3/04 (2023.01); G06N 3/08 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 16/65 (2019.01); G06N 3/04 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method for disambiguating musical artist names carried out by a computing system, the method comprising:
inputting to a multi-modal artificial neural network (ANN) implemented on the computing system a first plurality of musical-artist-release records (MARRs), each respective MARR record being associated with a given musical release of a given artist of a second plurality of musical artists, and each respective MARR record comprising (i) a respective release ID of the given musical release and a unique artist identifier (ID) of the given artist, and (ii) respective release data assigned according to release-data categories that comprise particular music media content associated with the given musical release, and a set of metadata categories including a sub-definitive musician name of the given artist and a subset of release subcategories specific to the given musical release;
forming each possible n-tuple grouping of distinct MARRs from among the first plurality;
for each respective n-tuple grouping, concurrently applying the multi-modal ANN to each respective MARR of the respective n-tuple to generate an associated, respective release feature vector (RFV), wherein for each respective MARR, the associated, respective RFV is also associated with the unique artist ID of the respective MARR, and comprises a set of sub-feature vectors, each sub-feature vector respectively characterizing release data of a different one of the release-data categories of the respective MARR, and wherein the respective RFVs associated with the respective MARRs are defined within a multi-dimensional RFV space;
for each respective n-tuple grouping, concurrently training the multi-modal ANN and adjusting the associated, respective RFVs to cluster in RFV space of all the respective RFVs associated with the same unique artist ID, and to maximally separate in RFV space the respective RFVs associated with different unique artist IDs; and
storing each respective MARR together with its adjusted, associated, respective RFV in a musical release database.