This paper investigates a novel neural scoring method, based on conventional i-vectors, to perform speaker diarization and linking of large collections of recordings. Using triplet loss for training, the network projects i-vectors in a space that better separates speakers in terms of cosine similarity. Experiments are run on two French TV collections built from REPERE [1] and ETAPE [2] campaigns corpora, the system being trained on French Radio data. Results indicate that the proposed approach outperforms conventional cosine and Probabilistic Linear Discriminant Analysis scoring methods on both within-and cross-recording diarization tasks, with a Diarization Error Rate reduction of 14% in average.

No abstract

This paper investigates iterative PLDA adaptation for crossshow speaker diarization applied to small collections of French TV archives based on an i-vector framework. Using the target collection itself for unsupervised adaptation, PLDA parameters are iteratively tuned while score normalization is applied for convergence. Performances are compared, using combinations of target and external data for training and adaptation. The experiments on two distinct target corpora show that the proposed framework can gradually improve an existing system trained on external annotated data. Such results indicate that performing speaker diarization on small collections of unlabeled audio archives should only rely on the availability of a sufficient bootstrap system, which can be incrementally adapted to every target collection. The proposed framework also widens the range of acceptable speaker clustering thresholds for a given performance objective.

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.

Biometrics authentication is now widely deployed, and from that omnipresence comes the necessity to protect private data. Recent studies proved touchscreen handwritten digits to be a reliable biometrics. We set a threat model based on that biometrics: in the event of theft of unlabelled embeddings of handwritten digits, we propose a labelling method inspired by recent unsupervised translation algorithms. Provided a set of unlabelled embeddings known to have been produced by a Long Short Term Memory Recurrent Neural Network (LSTM RNN), we demonstrate that inferring their labels is possible. The proposed approach involves label-wise clustering of the embeddings and label identification of each group by matching their distribution to the label-relative classes of a comparison hand-crafted labeled set of embeddings. Cluster labelling is done through a two steps process including a genetic algorithm that finds the N-best matching hypotheses before a fine-tuning of those N-candidates. The proposed method was able to infer the correct labels on 100 randomised runs on different dataset splits.