SecondHandSongs: Data Architecture and Research Applications
SecondHandSongs serves as a comprehensive community-driven repository dedicated to documenting cover versions of musical works. By leveraging the contributions of an active community, the platform has evolved from a niche archive into a massive database that tracks the relationship between original songs and their subsequent interpretations.
Growth and Scale of the Database
The scale of the SecondHandSongs database has expanded significantly over the last two decades. In 2007, the project contained approximately 60,000 covers. By 2020, the collection had grown substantially, reaching a milestone of one million covers.
[ไม่มีภาพประกอบ]
Data Schema and Identifiers
To maintain accuracy across millions of entries, SecondHandSongs employs a specific data schema—a structured framework for organizing information—using two primary identifiers:
- Work ID: A unique identifier assigned to the underlying musical composition.
- Performance ID: A unique identifier assigned to a specific version of a work, whether it is the original recording or a cover by a performer.
In this system, a "work" is treated as an equivalence class, which is essentially a curated list of all performances that share the same underlying song. Generally, the database records only one performance per performer for any given work.
Derived Datasets for Academic Research
The structured nature of SecondHandSongs has made it an invaluable resource for the scientific community, leading to the creation of several derived datasets used in music information retrieval.
The Million Song Dataset Intersection
In 2011, the Million Song Dataset (MSD) project released a subset that represented the intersection of SHS and MSD data. At the time of its release, this constituted the largest available dataset of cover songs for academic research purposes.
The SHS100k Benchmark
Furthering the utility for artificial intelligence, the SHS100k dataset was later released specifically for machine learning. This dataset contains 100,000 covers spanning 10,000 distinct works. It has since established itself as a primary benchmark for cover-song identification, the process of automatically recognizing a song regardless of the performer or arrangement.
Key Facts
- Database growth: Increased from 60,000 covers in 2007 to 1 million by 2020.
- Identification system: Uses Work IDs for compositions and Performance IDs for specific recordings.
- Academic impact: Provided the foundation for the SHS100k machine learning benchmark.
- SHS100k scale: Includes 100k covers of 10k unique works.
| Year/Dataset | Scale/Content | Primary Use |
|---|---|---|
| 2007 Database | 60,000 covers | Community Archive |
| 2020 Database | 1,000,000 covers | Community Archive |
| MSD Subset (2011) | SHS & MSD Intersection | Academic Research |
| SHS100k | 100k covers / 10k works | Machine Learning Benchmark |
Frequently Asked Questions
What is the difference between a Work ID and a Performance ID?
A Work ID identifies the original song composition itself, while a Performance ID identifies a specific recording of that song by a particular artist.
How has the size of the SecondHandSongs database changed?
The database grew from 60,000 covers in 2007 to one million covers by the year 2020.
What is an equivalence class in the context of SecondHandSongs?
An equivalence class is a list of various performances (covers and originals) that all belong to the same underlying musical work.
What is the SHS100k dataset used for?
The SHS100k dataset is used as a benchmark for machine learning, specifically for the task of cover-song identification.
How many works are represented in the SHS100k dataset?
The SHS100k dataset contains 100,000 covers derived from 10,000 unique works.