VoxCeleb multi-speaker speaker verification trials

Extension of the VoxCeleb1-O trial set for speaker verification to scenarios where also speech mixtures are considered.

Here, lists for the single speaker vs. mixture (s vs m) and mixture vs. mixture (m vs. m) are available.

Each trial consists of the following entries:

s vs. m:

(Label, IDa, [ID_b1, ID_b2], log_weights)

m vs m:

(Label, [ID_a1, ID_a2], [ID_b1, ID_b2], log_weights_a, log_weights_b)

The IDs are the VoxCeleb utterances used for mixing, and the log weights are the logarithmic scaling factors used to scale the respective sources before mixing.  

For more details, please refer to the following publication:

@inproceedings{cordlandwehr23_interspeech,
  title     = {A Teacher-Student Approach for Extracting Informative Speaker Embeddings From Speech Mixtures},
  author    = {Tobias Cord-Landwehr and Christoph Boeddeker and Cătălin Zorilă and Rama Doddipatla and Reinhold Haeb-Umbach},
  year      = {2023},
  booktitle = {Interspeech 2023},
  pages     = {4703--4707},
  doi       = {10.21437/Interspeech.2023-1379},
  issn      = {2958-1796},
}