Audio stem identification systems and methods
Methods, systems and computer program products are provided for determining acoustic feature vectors of query and target items in a first vector space, and mapping the acoustic feature vectors to a second vector space having a lower dimension. The distribution of vectors in the second vector space can then be used to identify items from the same songs, and/or items that are complementary. A mapping function is trained using a machine learning algorithm, such that complementary audio items are closer in the second vector space than the first, according to a given distance metric.
1. A stem identifier, comprising:
at least one processor;
at least one memory storing instructions which when executed by the at least one processor cause the at least one processor to:
receive a target input audio type;
receive a query vector corresponding to a query stem, wherein the query stem represents an instrumental stem or a vocal stem;
compare the query vector and a plurality of target vectors in a vector space, each of the plurality of target vectors corresponding to a plurality of stems;
generate a plurality of stem identifications, each stem identification including a stem identifier and a corresponding likelihood value indicating a probability a stem corresponding to the stem identifier is a match for the query stem and the target input audio type; and
output to an editor tool, based on the comparison, the plurality of stem identifications, enabling a user of the editor tool to assemble a group of stems including at least one of the plurality of stems corresponding to the plurality of stem identifications.
2. The stem identifier according to claim 1 , wherein the plurality of stem identifications is an ordered list of stem identifications.
3. The stem identifier according to claim 1 , wherein the at least one memory further stores instructions which when executed by the at least one processor cause the at least one processor to:
obtain the corresponding likelihood value by computing each distance between the query vector and each vector in the vector space.
4. The stem identifier according to claim 3 , wherein the at least one memory further stores instructions which when executed by the at least one processor cause the at least one processor to:
determine a predetermined number of vectors in the vector space closest to the query vector based on each computed distance.
5. The stem identifier according to claim 4 , wherein the at least one memory further stores instructions which when executed by the at least one processor cause the at least one processor to:
normalize the distance as: 1−D/max_distance,
where max_distance corresponds to a maximum possible distance in a given space using a given distance metric.
6. The stem identifier according to claim 5 , wherein the given space is an N-dimensional hypercube and the distance metric is a Euclidean distance, wherein N is an integer.
7. The stem identifier according to claim 1 , wherein the at least one memory further stores instructions which when executed by the at least one processor cause the at least one processor to:
iteratively train a mapping function to reduce a distance between a plurality of complementary stems and to increase a distance of non-complementary stems.
8. A method of identifying audio stems, comprising:
receiving a target input audio type;
receiving a query vector corresponding to a query stem, wherein the query stem represents an instrumental stem or a vocal stem;
comparing the query vector and a plurality of target vectors in a vector space, each of the plurality of target vectors corresponding to a plurality of stems;
generate a plurality of stem identifications, each stem identification including a stem identifier and a corresponding likelihood value indicating a probability a stem corresponding to the stem identifier is a match for the query stem and the target input audio type; and
outputting, to an editor tool, based on the comparison, the plurality of stem identifications, enabling a user of the editor tool to assemble a group of stems including at least one of the plurality of stems corresponding to the plurality of stem identifications.
9. The method according to claim 8 ,
wherein the plurality of stem identifications is an ordered list of stem identifications.
10. The method according to claim 8 , further comprising:
obtaining the corresponding likelihood value includes:
computing each distance between the query vector and each vector in the vector space.
11. The method according to claim 10 , further comprising:
determining a predetermined number of vectors in the vector space closest to the query vector based on each computed distance.
12. The method according to claim 11 , further comprising:
normalizing the distance as: 1−D/max_distance,
where max_distance corresponds to a maximum possible distance in a given space using a given distance metric.
13. The method according to claim 12 , wherein the given space is an N-dimensional hypercube and the distance metric is a Euclidean distance, wherein N is an integer.
14. The method according to claim 8 , further comprising:
iteratively training a mapping function to reduce a distance between a plurality of complementary stems and to increase a distance of non-complementary stems.