Acomus
Acoustic Covers of Music
The Acoustic Covers of Music (ACOMUS) dataset born to partially cover the lack of musical corpus for sung speech recognition. This corpus is composed of several acoustic covers versions of popular songs from YouTube.
The motivation of this project was to create a suitable corpus for my MSc Dissertation project on sung speech recognition. This dataset was designed and constructed for academic porpoises only, and all the songs and its lyrics are property of the creator artist, and all the credits belong to them.
This is a small dataset that can be used for evaluation and benchmarking.
Construction Process
I built this dataset following these steps:
- Video collection — Searched YouTube for hundreds of videos of people singing song covers accompanied by an acoustic guitar or piano.
- Lyrics retrieval — Searched online for the official lyrics of each song.
- Audio extraction and segmentation — Extracted the audio from each video and split it into complete lyric phrases.
- Lyrics alignment — Aligned the lyrics with the corresponding audio segments.
- Lyrics correction — Manually corrected any discrepancies between the lyrics and the words actually sung, which can occur when a singer deviates from the original text for artistic reasons.
Characteristics
The corpus was designed and constructed with the following specifications:
- The songs are mainly interpreted by Amateur Artists and around 10% are known artists.
- The dataset is separated in a balanced number of Male and Female artists.
- The interpretations have one accompaniment instrument, and just some few cases have more than one.
- The 80% of the database are accompanied by acoustic guitar and 20% by piano.
Table 1 summarizes the total and annotated duration of the corpus by accompaniment instrument:
Table 1: Total vs. annotated audio duration in ACOMUS, by instrument (guitar and piano).
| Instrument | Annotated | Total Time | Annotated Time |
|---|---|---|---|
| Guitar | 100 | 389 min. | 233 min. |
| Piano | 20 | 77 min. | 49 min. |
Table 2 shows how the singers and songs are distributed by sex, along with the split between guitar and piano accompaniment for each:
Table 2: Distribution of ACOMUS singers and songs by sex, with instrument breakdown (guitar vs. piano).
| Sex | Unique Singers | Total Songs | Guitar | Piano |
|---|---|---|---|---|
| Female | 86 | 115 | 91 | 24 |
| Male | 89 | 125 | 109 | 16 |
Download
Currently, only the code to generate the dataset is available on GitHub, and it is in Python 2.7. I am planning on updating the code and on releasing the audio excerpts to stop relaying on YouTube.