Phonetic Richness for Improved Automatic Speaker Verification: Aplawd-Based Speaker Verification Protocols

This data was created as part of the work entitled "Phonetic Richness for Improved Automatic Speaker Verification", published in EUSIPCO 2024.

In doing our experimental evaluation, Pindrop used FLAC files from the APLAWD Markings Dataset. You can obtain copies of these same FLAC files directly from the developers here: https://github.com/serwy/aplawdw

This data contains two automatic speaker verification protocols: "Aplawd" and "Aplawd-Repetitive". See below for details regarding the structure of the data for each of these protocols.

Files defining the Aplawd protocol:> Pindrop_aplawd_protocol/models.csv

This file defines what files should be used to create each enrollment model used in this protocol.

Columns:

enroll_id: unique string identifying the enrollment model that this row corresponds to

filename: filename specifying the file from the APLAWD Markings Dataset to be used in creating the enrollment model specified for this row by enroll_id.

Note: a single model is created for each enroll_id, using all of the files specified for the given enroll_id

> Pindrop_aplawd_protocol/trials_same_gender.csv

This file defines the trials that make up this protocol, specifying pairs of enrollment model and probe audios to test against eachother.

Columns:

enroll_id: unique string identifying the enrollment model whose identity will be compared against in the trial defined by this row.

probe_filename: filename specifying the file from the APLAWD Markings Dataset for which the identity of the speaker should be compared against the identity of the corresponding enrollment model for this row.

probe_identity: unique string identifying the true speaker identity for this row's probe.

Files defining the Aplawd-Repetitive protocol:

(NOTE: the key difference from the above Aplawd protocol is in the probe_filename column of the trials_same_gender.csv file)

> Pindrop_aplawd_repetitive_protocol/models.csv

This file defines what files should be used to create each enrollment model used in this protocol.

Columns:

enroll_id: unique string identifying the enrollment model that this row corresponds to

filename: filename specifying the file from the APLAWD Markings Dataset to be used in creating the enrollment model specified for this row by enroll_id.

 Note: a single model is created for each enroll_id, using all of the files specified for the given enroll_id

> Pindrop_aplawd_repetitive_protocol/trials_same_gender.csv

This file defines the trials that make up this protocol, specifying pairs of enrollment model and probe audios to test against eachother.

Columns:

enroll_id: unique string identifying the enrollment model whose identity will be compared against in the trial defined by this row.

probe_filename: an underscore ("_") delimited string of filename(s) specifying the file(s) from the APLAWD Markings Dataset that should be concatenated to form the probe audio for this row's trial. The resulting probe's speaker identity should be compared against the identity of the corresponding enrollment model for this row.

probe_identity: unique string identifying the true speaker identity for this row's probe.