Reproduce “Deep Speaker Feature Learning for Text-independent Speaker Verification” on OpenResearch

About this paper

Lantian Li, Yixiang Chen, Ying Shi, Zhiyuan Tang, Dong Wang

Researchers at Tsinghua University developed a Convolutional Time-delay Deep Neural Network (CT-DNN) to extract highly discriminative speaker features for text-independent speaker verification. This method demonstrated superior performance on short utterances, achieving a 2.15% Equal Error Rate (EER) for 3-second test utterances and effectively verifying speakers from segments as short as 0.3 seconds.

Key insights

• Speaker identity can be accurately extracted as a deterministic short-time property from very brief speech segments (e.g., 0.3 seconds), rather than solely relying on long-time statistical patterns.
• Well-designed deep neural networks can produce speaker features that are sufficiently strong and discriminative on their own, validating a 'pure feature learning' approach for speaker verification.
• Traditional statistical back-end models like PLDA, which assume Gaussian distributions, are generally unsuitable for highly abstracted deep speaker features (d-vectors) and can degrade performance.

Summary by alphaXiv.