NMF-based Improvement of DNN and LSTM Pre-Training for Speech Enhancemet

Q: How can I download an article?

To download an article from SID, first log in to the site, search for the article title, and click on the 'Download Article' option.

Q: How can I download an ISI article?

To download an ISI article on SID, enter the keyword or article title in the search bar, view the relevant results, click on the desired article, and select the 'Download Article' option.

Q: How can I access the SID database?

To access the SID database, visit SID.ir, create an account, and log in to access scientific resources.

Q: Is downloading articles from SID free?

Some articles on SID are available for free, while others require payment. Details are specified on the article's page.

Safari Dehnavi Razieh; Seyedin Sanaz

مرکز اطلاعات علمی Scientific Information Database (SID) - Trusted Source for Research and Academic Resources

Journal Paper

Paper Information

Journal: International Journal of Information and Communication Technology Research Year:2023 | Volume:15 | Issue:3 Page(s): 53-65

Download Full-Text

View:

Download:

Cites:

Information Journal Paper

Title

NMF-based Improvement of DNN and LSTM Pre-Training for Speech Enhancemet

Author(s)

Safari Dehnavi Razieh | Seyedin Sanaz | Issue Writer Certificate

Keywords

pre-training

deep neural networks (DNN)

long short-term memory (LSTM)

non-negative matrix factorization (NMF)

speech enhancement

basis matrix

noise classification

Abstract

A novel pre-training method is proposed to improve deep-neural-networks (DNN) and long-short-term-memory (LSTM) performance, and reduce the local minimum problem for speech enhancement. We propose initializing the last layer weights of DNN and LSTM by Non-Negative-Matrix-Factorization (NMF) basis transposed values instead of random weights. Due to its ability to extract speech features even in presence of non-stationary noises, NMF is faster and more successful than previous pre-training methods for network convergence. Using NMF basis matrix in the first layer along with another pre-training method is also proposed. To achieve better results, we further propose training individual models for each noise type based on a noise classification strategy. The evaluation of the proposed method on TIMIT data shows that it outperforms the baselines significantly in terms of perceptual-evaluation-of-speech-quality (PESQ) and other objective measures. Our method outperforms the baselines in terms of PESQ up to 0.17, with an improvement percentage of 3.4%.

Multimedia

No record.

Cites

No record.

References

No record.

Cite

Related Journal Papers

No record.

Related Seminar Papers

No record.

Related Plans