The Speakers in the Wild (SITW) speaker recognition database contains hand-annotated speech samples from open-source media for the purpose of benchmarking text-independent speaker recognition technology on single and multi-speaker audio acquired across unconstrained or "wild" conditions. The database consists of recordings of 299 speakers, with an average of eight different sessions per person. Unlike existing databases for speaker recognition, this data was not collected under controlled conditions and thus contains real noise, reverberation, intraspeaker variability and compression artifacts. These factors are often convolved in the real world, as the SITW data shows, and they make SITW a challenging database for single-and multispeaker recognition

With the recent introduction of speaker embeddings for text-independent speaker recognition, many fundamental questions require addressing in order to fast-track the development of this new era of technology. Of particular interest is the ability of the speaker embeddings network to leverage artificially degraded data at a far greater rate beyond prior technologies, even in the evaluation of naturally degraded data. In this study, we aim to explore some of the fundamental requirements for building a good speaker embeddings extractor. We analyze the impact of voice activity detection, types of degradation, the amount of degraded data, and number of speakers required for a good network. These aspects are analyzed over a large set of 11 conditions from 7 evaluation datasets. We lay out a set of recommendations for training the network based on the observed trends. By applying these recommendations to enhance the default recipe provided in the Kaldi toolkit, a significant gain of 13-21% on the Speakers in the Wild and NIST SRE'16 datasets is achieved.

No abstract

The output scores of most speaker recognition systems are not directly interpretable as stand-alone values. For this reason, a calibration step is usually performed on the scores to convert them into proper likelihood ratios (LR), which have a clear probabilistic interpretation. The standard calibration approach transforms the system scores using a linear function trained using data selected to closely match the evaluation conditions. This selection, though, is not feasible when the evaluation conditions are unknown. In previous work, we proposed a calibration approach for this scenario called trialbased calibration (TBC). TBC trains a separate calibration model for each test trial using data that is dynamically selected from a candidate training set to match the conditions of the trial. In this work, we extend the TBC method, proposing (1) a new similarity metric for selecting training data that results in significant gains over the one proposed in the original work, (2) a new option that enables the system to reject a trial when not enough matched data is available for training the calibration model, and (3) the use of regularization to improve the robustness of the calibration models trained for each trial. We test the proposed algorithms on a development set composed of several conditions and on the FBI multi-condition speaker recognition dataset, and we demonstrate that the proposed approach reduces calibration loss to values close to 0 for most conditions when matched calibration data is available for selection and that it can reject most trials for which relevant calibration data is unavailable.

This paper studies a novel audio segmentation-by-classification approach based on factor analysis. The proposed technique compensates the within-class variability by using class-dependent factor loading matrices and obtains the scores by computing the log-likelihood ratio for the class model to a non-class model over fixed-length windows. Afterwards, these scores are smoothed to yield longer contiguous segments of the same class by means of different back-end systems. Unlike previous solutions, our proposal does not make use of specific acoustic features and does not need a hierarchical structure. The proposed method is applied to segment and classify audios coming from TV shows into five different acoustic classes: speech, music, speech with music, speech with noise, and others. The technique is compared to a hierarchical system with specific acoustic features achieving a significant error reduction.

Audio segmentation is important as a pre-processing task to improve the performance of many speech technology tasks and, therefore, it has an undoubted research interest. This paper describes the database, the metric, the systems and the results for the Albayzín-2014 audio segmentation campaign. In contrast to previous evaluations where the task was the segmentation of non-overlapping classes, Albayzín-2014 evaluation proposes the delimitation of the presence of speech, music and/or noise that can be found simultaneously. The database used in the evaluation was created by fusing different media and noises in order to increase the difficulty of the task. Seven segmentation systems from four different research groups were evaluated and combined. Their experimental results were analyzed and compared with the aim of providing a benchmark and showing up the promising directions in this field.

In this paper, we analyze and assess the impact of critical metadata factors on the calibration performance of speaker recognition systems. In particular, we study the effect of duration, distance, language, and gender by using a variety of datasets and systematically varying the conditions in the evaluation and calibration sets. For all experiments, the system is based on ivectors and a probabilistic linear discriminant analysis (PLDA) back-end and linear calibration. We measure system performance in terms of calibration loss. Our experiments reveal (i) a large degradation when the duration used for calibration is significantly different from that in the evaluation set; (ii) no significant degradation when a different gender is used for calibration than for evaluation; (iii) a large degradation when microphone distance is significantly different between the sets; and (iv) a small loss for closely related languages and languages with shared vocabulary. This analysis will be beneficial in the development of speaker recognition systems for use in unseen environments and for forensic speaker recognition analysts when selecting relevant population data.

Multi-domain language recognition involves the application of a language identification (LID) system to identify languages in more than one domain. This problem was the focus of the recent NIST LRE 2017, and this article presents the findings from the SRI team during system development for the evaluation. Approaches found to provide robustness in multi-domain LID include a domain-and-language-weighted Gaussian backend classifier, duration-aware calibration, and a source normalized multi-resolution neural network backend. The recently developed speaker embeddings technology is also applied to the task of language recognition, showing great potential for future LID research.