GMM_UBM_loop: Performs GMM-UBM-based LR calculation on a set of speakers

In justinjhlo/fvclrr: Likelihood Ratio Calculation and Testing in Forensic Voice Comparison

Performs GMM-UBM-based LR calculation on a set of speakers

Description

Models reference population using a single GMM (UBM) and models suspect by adapting from UBM. Called by LR_test when the mode GMM-UBM is chosen. Procedure is based on Reynolds, D.A., Quatieri, T.F. & Dunn, R.B. (2000). Speaker Verification Using Gaussian Mixture Models. Digital Signal Processing, 10(1-3), 19-41.

Usage

GMM_UBM_loop(
  bg_data,
  by_speaker_data,
  test_speakers,
  G = 8,
  r = 16,
  background_model = NULL
)

Arguments

bg_data	A data frame of background data. The first column identifies speakers. All other columns contain data.
by_speaker_data	A named list of 2 sub-lists: suspect_data and offender_data. Contains data in the test_speakers set, divided by speaker. Each is a named list of data frames, with speaker IDs as the names.
test_speakers	A vector of test speakers.
G	Number of components in the Gaussian Mixture Model.
r	Relevance factor for speaker adaptation.
background_model	(Optional) Model pre-fitted on bg_data. Useful when the same model is used for different test data sets.

Value

A named list of 3 items:

• likelihood_ratio_matrix: A data frame. Rows and columns are named after the speaker identifiers. Each row and column represents a speaker as suspect and offender respectively, and each cell contains a single logLR score.

• cllr: Numeric. Reports the logLR cost.

• eer: Numeric. Reports the equal error rate (between 0 and 1).

justinjhlo/fvclrr documentation built on June 14, 2025, 5:43 a.m.