Search

This work addresses the cross-corpora generalization issue for the low-resourced spoken language identification (LID) problem. We have conducted the experiments in the context of Indian LID and identified strikingly poor cross-corpora generalization due to corpora-dependent non-lingual biases. Our contribution to this work is twofold. First, we propose domain diversification, which diversifies the limited training data using different audio data augmentation methods. We then propose the concept of maximally diversity-aware cascaded augmentations and optimize the augmentation fold-factor for effective diversification of the training data. Second, we introduce the idea of domain generalization considering the augmentation methods as pseudo-domains. Towards this, we investigate both domain-invariant and domain-aware approaches. Our LID system is based on the state-of-the-art emphasized channel attention, propagation, and aggregation based time delay neural network (ECAPA-TDNN) architecture. We have conducted extensive experiments with three widely used corpora for Indian LID research. In addition, we conduct a final blind evaluation of our proposed methods on the Indian subset of VoxLingua107 corpus collected in the wild. Our experiments demonstrate that the proposed domain diversification is more promising over commonly used simple augmentation methods. The study also reveals that domain generalization is a more effective solution than domain diversification. We also notice that domain-aware learning performs better for same-corpora LID, whereas domain-invariant learning is more suitable for cross-corpora generalization. Compared to basic ECAPA-TDNN, its proposed domain-invariant extensions improve the cross-corpora EER up to 5.23%. In contrast, the proposed domain-aware extensions also improve performance for same-corpora test scenarios., Accepted for publication in Elsevier Computer Speech & Language

Standard Mel frequency cepstrum coefficient (MFCC) computation technique utilizes discrete cosine transform (DCT) for decorrelating log energies of filter bank output. The use of DCT is reasonable here as the covariance matrix of Mel filter bank log energy (MFLE) can be compared with that of highly correlated Markov-I process. This full-band based MFCC computation technique where each of the filter bank output has contribution to all coefficients, has two main disadvantages. First, the covariance matrix of the log energies does not exactly follow Markov-I property. Second, full-band based MFCC feature gets severely degraded when speech signal is corrupted with narrow-band channel noise, though few filter bank outputs may remain unaffected. In this work, we have studied a class of linear transformation techniques based on block wise transformation of MFLE which effectively decorrelate the filter bank log energies and also capture speech information in an efficient manner. A thorough study has been carried out on the block based transformation approach by investigating a new partitioning technique that highlights associated advantages. This article also reports a novel feature extraction scheme which captures complementary information to wide band information; that otherwise remains undetected by standard MFCC and proposed block transform (BT) techniques. The proposed features are evaluated on NIST SRE databases using Gaussian mixture model-universal background model (GMM-UBM) based speaker recognition system. We have obtained significant performance improvement over baseline features for both matched and mismatched condition, also for standard and narrow-band noises. The proposed method achieves significant performance improvement in presence of narrow-band noise when clubbed with missing feature theory based score computation scheme.