[Paper Review] Autoencoder based Domain Adaptation for Speaker Recognition under Insufficient Channel Information

Suwon Shon, Seongkyu Mun+2 more|arXiv (Cornell University)|Aug 3, 2017

Speech Recognition and Synthesis21 references3 citations

TL;DR

This paper proposes Autoencoder-based Domain Adaptation (AEDA), an unsupervised method that combines standard and denoising autoencoders to adapt speaker recognition systems from resource-rich development sets to low-resource test domains with insufficient channel information. Evaluated on the Interspeech 2017 Domain Adaptation Challenge, AEDA achieves significant performance gains over baselines and prior methods under domain mismatched conditions.

ABSTRACT

In real-life conditions, mismatch between development and test domain degrades speaker recognition performance. To solve the issue, many researchers explored domain adaptation approaches using matched in-domain dataset. However, adaptation would be not effective if the dataset is insufficient to estimate channel variability of the domain. In this paper, we explore the problem of performance degradation under such a situation of insufficient channel information. In order to exploit limited in-domain dataset effectively, we propose an unsupervised domain adaptation approach using Autoencoder based Domain Adaptation (AEDA). The proposed approach combines an autoencoder with a denoising autoencoder to adapt resource-rich development dataset to test domain. The proposed technique is evaluated on the Domain Adaptation Challenge 13 experimental protocols that is widely used in speaker recognition for domain mismatched condition. The results show significant improvements over baselines and results from other prior studies.

Motivation & Objective

• To address performance degradation in speaker recognition when test domain channel characteristics are poorly represented in the training data.
• To develop an unsupervised domain adaptation method that effectively utilizes limited in-domain data for domain shift mitigation.
• To improve generalization of speaker recognition systems in real-world conditions with insufficient channel variability information.
• To explore the effectiveness of combining autoencoders and denoising autoencoders for domain adaptation in low-resource domain scenarios.

Proposed method

• The method employs a standard autoencoder to learn a compact, shared representation across development and test domains.
• A denoising autoencoder is used to enhance robustness by reconstructing clean inputs from corrupted versions, improving generalization to unseen channel conditions.
• The two autoencoders are jointly trained to align the latent space of the resource-rich development data with the limited in-domain test data.
• Feature-level domain adaptation is performed by projecting development set features into a shared latent space that reduces domain shift.
• The approach does not require in-domain labels, making it suitable for unsupervised adaptation.
• The final speaker recognition system is trained on the adapted features, leveraging the adapted representation for improved generalization.

Experimental results

Research questions

• RQ1 Can an unsupervised domain adaptation method effectively improve speaker recognition performance when only limited in-domain data is available?
• RQ2 How does combining standard and denoising autoencoders enhance robustness to domain shift in speaker recognition?
• RQ3 To what extent does the proposed AEDA method reduce performance degradation under domain mismatch compared to existing baselines?
• RQ4 Does the joint use of autoencoders and denoising autoencoders lead to better generalization in low-resource channel conditions?

Key findings

• The proposed AEDA method achieves significant performance improvements over baseline systems on the Domain Adaptation Challenge 2013 protocol.
• AEDA outperforms prior state-of-the-art methods in speaker recognition under domain mismatched conditions with limited in-domain data.
• The combination of standard and denoising autoencoders leads to more robust feature representations under channel variability.
• The unsupervised nature of the method enables effective adaptation without requiring labeled in-domain data.