Towards A Unified Conformer Structure: from ASR to ASV Task (2211.07201v2)

Published 14 Nov 2022 in eess.AS and cs.SD

Abstract: Transformer has achieved extraordinary performance in Natural Language Processing and Computer Vision tasks thanks to its powerful self-attention mechanism, and its variant Conformer has become a state-of-the-art architecture in the field of Automatic Speech Recognition (ASR). However, the main-stream architecture for Automatic Speaker Verification (ASV) is convolutional Neural Networks, and there is still much room for research on the Conformer based ASV. In this paper, firstly, we modify the Conformer architecture from ASR to ASV with very minor changes. Length-Scaled Attention (LSA) method and Sharpness-Aware Minimizationis (SAM) are adopted to improve model generalization. Experiments conducted on VoxCeleb and CN-Celeb show that our Conformer based ASV achieves competitive performance compared with the popular ECAPA-TDNN. Secondly, inspired by the transfer learning strategy, ASV Conformer is natural to be initialized from the pretrained ASR model. Via parameter transferring, self-attention mechanism could better focus on the relationship between sequence features, brings about 11% relative improvement in EER on test set of VoxCeleb and CN-Celeb, which reveals the potential of Conformer to unify ASV and ASR task. Finally, we provide a runtime in ASV-Subtools to evaluate its inference speed in production scenario. Our code is released at https://github.com/Snowdar/asv-subtools/tree/master/doc/papers/conformer.md.

Authors (5)

Dexin Liao (2 papers)
Tao Jiang (274 papers)
Feng Wang (408 papers)
Lin Li (329 papers)
Qingyang Hong (29 papers)

Citations (9)

View on Semantic Scholar

Summary

An Evaluation of the Unified Conformer Structure for ASR and ASV Tasks

The paper "Towards a Unified Conformer Structure: From ASR to ASV Task" provides a comprehensive analysis of adapting the Conformer architecture, traditionally successful in Automatic Speech Recognition (ASR), to the domain of Automatic Speaker Verification (ASV). The paper ventures into exploring the applicability and performance of the Conformer model, aiming for unified architecture across both ASR and ASV tasks.

Key Contributions and Methods

The authors propose modifications to the Conformer architecture with minimal changes to transition it from ASR to ASV. They introduce two novel methods to enhance model generalization: Length-Scaled Attention (LSA) and Sharpness-Aware Minimization (SAM). LSA is designed to stabilize attention weight distribution for inputs of varying lengths, while SAM aims to enhance robustness by seeking flatter minima in the loss landscape.

Additionally, the paper explores a parameter transferring strategy, leveraging pretrained ASR models as initializations for ASV tasks. This transfer approach aims to enhance the attention mechanism's focus on sequence features, resulting in an 11% relative improvement in Equal Error Rate (EER) on test sets from VoxCeleb and CN-Celeb.

Experimental Insights

The experiments target two major datasets: CN-Celeb and VoxCeleb. The ASV Conformer outperforms established networks like ECAPA-TDNN, demonstrating competitive results across multiple configurations. When employing LSA and SAM, improvements in generalization and reduction in EER are observed.

ASR transferring shows significant promise, particularly when incorporating Multi-CN and WenetSpeech datasets for pretraining. The methodology enhances ASV performance, suggesting the Conformer's capability to unify tasks, leading to potential advancements in multitask and multimodal machine learning frameworks.

Implications and Future Directions

This research opens up pathways for the convergence of ASR and ASV systems into a unified model, offering efficient resource utilization and potential for integrated audio processing solutions. The insights gained from applying transfer learning to these tasks may guide future developments in voice-related AI technologies.

Future research could benefit from further exploration of the underlying relationships between ASR and ASV tasks, aiming to solidify the theoretical foundations of this integration. Additionally, as technological demands evolve, there could be significant interest in investigating more sophisticated attention mechanisms and optimizing inference speed for real-time applications.

Conclusion

The paper presents substantial advancements in extending the Conformer architecture to ASV tasks, backed by rigorous experimental validation. The introduction of LSA, SAM, and parameter transferring highlights significant potential for future research in developing unified architectures applicable across diverse audio processing tasks.

PDF Markdown

Related Papers

Find Related Papers