For instruments? Instrument Dataset Recording Specifications

Record

  1. Record song by song for 60 mins in total.
  2. Dry vocal without reverb, delay, or instrumentals.
  3. Each song should be in only one language.
  4. If any two parts are exactly the same (both lyrics&melody), you should skip the second one.
  5. No speech in the dataset.
    1. for singing datasets, please don’t do non-melodic rap.
    2. for rapping datasets, please don’t do singing.
  6. No vocal overlaps or harmonies.
  7. No background noise or big room reflections.
  8. No obvious instrumental leaks from your headphones. (Try to lower your headphone volume)
  9. When two clips are connected, use cross-fade and do not cover any words, cross-fade over silence or breath or consonant only. No need to remove breaths.
  10. There should be at least 1s-long silence spaces at the beginning and end of each track.
  11. Personal style and expression are more important than accurate pitch or correct lyrics, missing the tempo occasionally is also acceptable.

image.png

image.png

image.png

Record for specific controls

In the 60mins of datasets, split it into the following parts with specific controls:

<aside>

General Performances

40 mins of the dataset —— record in the normal way with rich expressiveness, dynamic range, and key range.

</aside>

<aside>

Powerful Performances

10 mins of the dataset —— record specifically with rich power, make it sounds aggressive. Even for verse, sing in a very powerful, aggressive, angry way.

</aside>

<aside>

Soft Performances

10 mins of the dataset —— record specifically with weak strength, make it sounds soft and gentle than usual.

</aside>