First investigations on self trained speaker diarization

Abstract

This paper investigates self trained cross-show speaker diarization applied to collections of French TV archives, based on an i-vector/PLDA framework. The parameters used for i-vectors extraction and PLDA scoring are trained in a unsupervised way, using the data of the collection itself. Performances are compared, using combinations of target data and external data for training. The experimental results on two distinct target corpora show that using data from the corpora themselves to perform unsupervised iterative training and domain adaptation of PLDA parameters can improve an existing system, trained on external annotated data. Such results indicate that performing speaker indexation on small collections of unlabeled audio archives should only rely on the availability of a sufficient external corpus, which can be specifically adapted to every target collection. We show that a minimum collection size is required to exclude the use of such an external bootstrap.

Iterative PLDA Adaptation for Speaker Diarization

“…When α is too small, the DER quickly reaches a plateau and the system stops improving after one iteration. This was observed in [11], where adaptation was performed with concatenation of train and target data. This is equivalent to setting α as the ratio between the two dataset, in this case, close to 0.1.…”

“…In this paper, the target collections are too small to be solely used for unsupervised estimation of PLDA parameters [11], the use of a sufficient external labeled train corpus is mandatory to estimate those parameters. Figure 1 represents the overview of the diarization framework, including the two possible strategies for the training process: the bootstrap system (baseline) is represented with blue plain lines while the blue dashed lines correspond to the proposed adapted system.…”

“…Numerical details about the two corpora are presented in table 1. As opposed to [11], where the same shows are selected to build the target corpora, we decided to restrict our experiments to the annotated segments, in order to accurately evaluate the unsupervised part of the framework.…”

“…In our previous work [2,11], it was shown that an unsupervised diarization system, trained on a large multi-speaker unsegmented dataset, performs as well as a supervised one, indicating that annotations are not mandatory for training. However we noted that if a target dataset, used as training material, is too small, an effective unsupervised PLDA system cannot be trained.…”

“…We investigate the use of data from the target collection itself to perform iterative adaptation of PLDA parameters for speaker diarization, using manually annotated external data for bootstrap. The main difference with the system proposed in [11] is the refined adaptation step, the focus on the iterative process and the fact that only the cross-recording speaker linking is rerun through iterations, instead of the whole process.…”