From cslt Wiki
Revision as of 12:12, 21 April 2020 by Cslt (Talk | contribs)

Jump to: navigation, search


  • Back to 2017, we set our goal of deep speech factorizatoin. The first paper is published on ICASSP 2018
  • Lantian Li, Dong Wang, Yixiang Chen, Ying Shi, Zhiyuan Tang, "DEEP FACTORIZATION FOR SPEECH SIGNAL", ICASSP 2018. [1]
  • We noticed the problem of soft-max based training, due to the discardxing of the output layers


  • 2018/12/26, propose the idea of deep statistical speaker representation. That was based on VAE [3]


  • We noticed the impact of irregulation of deep speaker vectors, and tried to present normalization approaches
  • Yang Zhang and Lantian Li and Dong Wang, VAE-based regularization for deep speaker embedding, Interspeech 2019. [4]
  • 2019/04/20, "Normalization in speaker embedding", Speaker recognition workshop, Kunshan, Shanghai, [5]
  • 2019/07/17, Deep Feature Learning and Normalization for Speaker Recognition, report in India summr school [6]
  • 2019/08/14, present the first proposal that uses flow to model deep speaker featrues. (Report in Huawei group discussion)
  • 2019/10/27, present the initial idea of using flow to perform factorization, CSLT weekly meeting [7]


  • 2019/11/12, Yunqi start to work on DNF, using the subspace of the dimension to discriminte speakers [cvss 714]
  • 2019/12/20, I start to work on NF with constraint training. More understanding acheived for LDA. [cvss 741]
  • 2020/1/23, I noticed a bug in the DNF code, where the residual space was infact trained, so it is not a true dim-split DNF we hoped. [cvss 741]
  • 2020/1/27, I conjectured the normalization role of DNF, and informed YQ to perform a full-space experiment. The results are good. [8] [9]
  • 2020/1/28, I confirmed the normalization role of LDA for x-vectors. This forms the basic argument for the deep norm paper. [cvss 741]
  • 2020/2/10, Dong Wang, Deep Generative Models for Discriminative Tasks, CSLT weekly meeting. Present DNF
  • 2020/2/18, Deep norm paper submitted to IEEE Transactions.
  • 2020/2/24, I start working on optimal scoring for SRE, and establish the NL theory. The paper was submitted on 3.17 to APSIPA transaction.
  • 2020/3/21, I coined the NDA model, and completed the verification in 3 hours. This model can be used for scoring.
  • 2020/3/25, I designed the VAE-NF model, using NF to perform the generation net in VAE. It can generate more informative latent codes but the theory is not completed.
  • 2020/3/27, I extend the NDA to neural linear Gaussian model.
  • 2020/3/28, I extend the NDA to neural Bayesian model.
  • 2020/4/01, I started to work on NPCA. A simple L2 constraind algorithm was obtained, but this seems cannot find flexible manifold. Stop the work on 04/04.
  • 2020/4/04, Lantian demonstrated the NDA model with x-vector
  • 2020/4/06, I proposed the VAE-like NDA
  • 2020/4/08, Yunqi demonstrated VAE-like NDA with x-vector
  • 2020/4/10, I found the convergnece of the NDA model in simulation test (convergence of SB)
  • 2020/4/12, I proposed the Bayesian denoising approach
  • 2020/4/18, Lantian demonstrated the convergence of the NDA model with x vector
  • 2020/4/20, Zhiyuan demonstrated the capability of Bayesian denoising with white noise
  • 2020/4/21, I proposed to use Bayesian denoising with noise NF, and the architecture for speech separation based on the Bayesian inference.