turksfeat

Resources for Simple Speech Representation Learning from Perceptual Data.

This database contains 3200 pairs (2700 for training and 500 for testing) of phonemes whose perceived similarity is rated by workers on Amazon Mechnical Turk. Each pair is rated by 5 workers on a 1-to-5 discrete opinion scale.

There is also a smaller database (screening set) with only 61 token pairs but each rated by 25 workers.

Rating	Category	Description
5	Identical	they sound the same
4	Very Similar	almost the same but distinguishable
3	Fair	similar but not the same
2	Poor	not so similar but related
1	Bad	completely different sounds

Get VCTK

To use this database, you first need a copy of the VCTK speech corpus. Note that different license terms apply.

Data Format

• vctk-pairs-screen.csv, vctk-pairs-test.csv, vctk-pairs-train.csv

  speaker/speaker_utt1, time_start, time_end, speaker/speaker_utt2, time_start, time_end

  Example

  p225/p225_029,1.115,1.2,p225/p225_266,3.72,3.805
  p225/p225_301,1.725,1.825,p225/p225_165,3.77,3.875
  p226/p226_249,1.805,1.905,p226/p226_224,2.545,2.615
  p226/p226_249,1.545,1.65,p226/p226_144,3.495,3.625
  p226/p226_106,3.075,3.225,p226/p226_317,1.84,1.955
  

• vctk-ratings-screen.csv, vctk-ratings-test.csv, vctk-ratings-train.csv

  average, ratings[0], ratings[1], ratings[2], ...

  Example

  3.0,3,1,4,2,5
  3.6,2,5,3,5,3
  4.0,5,3,5,5,2
  4.6,5,5,4,4,5
  5.0,5,5,5,5,5
  

License

This database ("turksfeat") is made available under the Open Database License: http://opendatacommons.org/licenses/odbl/1.0/. Any rights in individual contents of the database are licensed under the Database Contents License: http://opendatacommons.org/licenses/dbcl/1.0/.

In short: attribution and share-alike.