[Paper Review] Cross-Corpora Spoken Language Identification with Domain Diversification and Generalization

Spandan Dey, Md Sahidullah+1 more|arXiv (Cornell University)|Feb 10, 2023

Speech Recognition and SynthesisComputer Science109 references14 citations

TL;DR

This paper proposes domain diversification and domain generalization techniques to improve cross-corpora spoken language identification (LID) in low-resource Indian languages. By using maximally diversity-aware cascaded audio augmentations and treating augmentation types as pseudo-domains, the authors enhance ECAPA-TDNN models via domain-invariant and domain-aware learning, reducing cross-corpora EER by up to 5.23% compared to baseline.

ABSTRACT

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.

Motivation & Objective

• Address the poor cross-corpora generalization of low-resource LID systems due to corpora-dependent non-lingual biases.
• Improve robustness of LID models trained on limited, in-house corpora when evaluated on unseen, diverse test corpora.
• Develop effective data augmentation strategies that simulate real-world domain variations to enhance training data diversity.
• Introduce domain generalization using augmentation-based pseudo-domains to improve generalization without requiring target-domain data.
• Compare domain-invariant and domain-aware learning strategies for optimal performance in same-corpora vs. cross-corpora settings.

Proposed method

• Propose domain diversification using multiple audio augmentation methods (e.g., speech enhancement, codec-based, noise injection) to increase training data variability.
• Introduce maximally diversity-aware cascaded augmentations with an optimized fold-factor to balance diversity and model performance.
• Frame augmentation types as pseudo-domains to enable domain generalization (DG) in the absence of real target-domain data.
• Implement domain-invariant learning using gradient reversal and MMD-based loss to minimize domain-specific representations.
• Implement domain-aware learning via multitask learning to allow the model to learn both language and domain-specific cues.
• Train and evaluate on ECAPA-TDNN, a state-of-the-art DNN architecture for speaker and language recognition, with extensions for DG.

Experimental results

Research questions

• RQ1 Can audio data augmentation strategies effectively simulate unseen domain variations to improve cross-corpora LID generalization?
• RQ2 Does treating augmentation types as pseudo-domains enable effective domain generalization in low-resource LID?
• RQ3 Is domain-invariant learning more effective than domain-aware learning for cross-corpora LID generalization?
• RQ4 How does domain diversification compare to standard augmentation in blind evaluation on unseen corpora?
• RQ5 Can domain generalization techniques reduce the performance gap between same-corpora and cross-corpora evaluations?

Key findings

• The proposed domain diversification with maximally diversity-aware cascaded augmentations outperforms conventional augmentation methods in blind evaluation on the Indian subset of VoxLingua107.
• Domain generalization significantly improves cross-corpora generalization, reducing the EER by up to 5.23% compared to the baseline ECAPA-TDNN.
• Domain-invariant learning is more effective for cross-corpora generalization, while domain-aware learning achieves better performance on same-corpora test sets.
• The use of pseudo-domains derived from diverse augmentation methods enables effective domain generalization without requiring labeled target-domain data.
• Audio augmentations such as speech enhancement and codec-based methods are particularly effective for improving robustness on KGP and LDC test sets.
• Despite improvements, a notable performance gap remains between same-corpora and cross-corpora evaluations, indicating room for further advancement with more realistic data augmentation.