End-to-End DNN Based Speaker Recognition Inspired by I-Vector and PLDA

Abstract

Recently, several end-to-end speaker verification systems based on deep neural networks (DNNs) have been proposed. These systems have been proven to be competitive for text-dependent tasks as well as for text-independent tasks with short utterances. However, for text-independent tasks with longer utterances, end-to-end systems are still outperformed by standard i-vector + PLDA systems. In this work, we develop an end-to-end speaker verification system that is initialized to mimic an i-vector + PLDA baseline. The system is then further trained in an end-to-end manner but regularized so that it does not deviate too far from the initial system. In this way we mitigate overfitting which normally limits the performance of endto-end systems. The proposed system outperforms the i-vector + PLDA baseline on both long and short duration utterances.

A Discriminative Condition-Aware Backend for Speaker Verification

“…As we will see in the results section, the approach above leads to good discrimination performance over a large set of conditions, sometimes improving over the baseline, in agreement with results found in [6]. Nevertheless, calibration performance of this approach is far from optimal on many conditions.…”

“…We propose a backend with the same functional form as the PLDAbackend explained in the previous section, but where all parameters are optimized jointly, in a manner similar to the one used in [6] (though, note that in this paper we only optimize jointly up to the backend stage instead of the full pipeline, as in Rohdin's paper). We first initialize all parameters in Equations (1), (2) and (5) as in the standard PLDA-based backend.…”

“…The parameters for the embedding extractor and the scorer are trained jointly to optimize binary cross-entropy. In [6], the authors propose to use an architecture that mimics the previous i-vector [8] pipeline for speaker verification, pretraining all its parameters separately and then fine tuning the full model to minimize binary cross-entropy. While these two approaches have the potential to result in well-calibrated scores, neither of the two papers show overall system performance, only discrimination performance.…”

Factorization of Discriminatively Trained i-Vector Extractor for Speaker Recognition

“…We show that, with such an approach, we can achieve a reasonable performance. Our results are perhaps not as competitive as those achieved with current state-of-the-art x-vector systems [18], nevertheless, we are now closer to our goal which is to further use this model in the fully end-to-end discriminative system [19] that can be initialized from a robust generative baseline. Figure 1: Scheme of an end-to-end speaker verification system based on a feed forward NN designed to mimic a generic speaker verification system ( [19]).…”

“…In [19], we had built an end-to-end system ( Fig. 1) that already seemingly fits our goal, but it was exactly the i-vector extractor component that posed the biggest challenge and we had to resort to ad-hoc simplifications, such as PCA-based dimensionality reduction of large dimensional sufficient statistics coming from the GMM-UBM.…”