Ziyue Zhao

Contributions to Neural Network-Based Speech Processing: Nonlinear Speech Prediction, Decoder Postprocessing, and Perceptual Loss Functions


Front	Back

ISBN:

978-3-8440-8779-6

Series:

Mitteilungen aus dem Institut für Nachrichtentechnik der Technischen Universität Braunschweig
Herausgeber: Prof. Dr.-Ing. U. Reimers, Prof. Dr.-Ing. T. Kürner and Prof. Dr.-Ing. T. Fingscheidt
Braunschweig

Volume:

Keywords:

Speech Coding; Speech Enhancement; Neural Networks

Type of publication:

Thesis

Language:

English

Pages:

154 pages

Figures:

28 figures

Weight:

228 g

Format:

21 x 14,8 cm

Binding:

Paperback

Price:

48,80 € / 61,10 SFr

Published:

October 2022

Buy:

Download:

Available PDF-Files for this title:

You need the Adobe Reader, to open the files. Here you get help and information, for the download.

These files are not printable.


	Document		Document
	Type		PDF
	Costs		36,60 EUR
	Action		Purchase in obligation and display of file - 1,1 MB (1198555 Byte)
	Action		Purchase in obligation and download of file - 1,1 MB (1198555 Byte)


	Document		Table of contents
	Type		PDF
	Costs		free
	Action		Display of file - 256 kB (261802 Byte)
	Action		Download of file - 256 kB (261802 Byte)

User settings for registered users

You can change your address here or download your paid documents again.

User:	Not logged in.
Actions:	Login / Register Forgotten your password?

Recommendation:

You want to recommend this title?

Review copy:

Here you can order a review copy.

Link:

You want to link this page? Click here.

Export citations:

Text
BibTex
RIS

Abstract:

Speech processing technologies are omnipresent in our daily communication products and services. Neural networks, as powerful data-driven models, have shown promising performance in various research fields, including speech processing. This thesis focuses on neural network-based speech processing, and it can be divided into three parts as follows.

In the field of speech prediction, a nonlinear speech predictor using the echo state network (ESN) is proposed as a novel adaptive prediction approach. This proposed nonlinear predictor shows better prediction performance than all baseline prediction methods in the simulations, including a predictor based on a long short-term memory (LSTM) structure. Second, the field of neural network-based speech enhancement puts focus on loss functions. A novel perceptual weighting filter (PWF) loss function motivated by the weighting filter from code-excited linear prediction (CELP) speech coding is proposed. A fully connected neural network (FCNN) and a convolutional neural network (CNN) are both used to evaluate the proposed loss functions, and the simulation results show their superior performance compared to baselines. Finally, neural network-based postprocessing for the enhancement of coded speech is studied. CNN-based postprocessors are proposed either to directly enhance the raw waveform in an end-to-end fashion, or to enhance the cepstral domain features using analysis synthesis. Furthermore, an advanced network structure, the fully convolutional recurrent network (FCRN), is utilized to enhance coded speech in the frequency domain, with the PWF loss function advantageously applied. The experimental results confirm the effectiveness of the proposed postprocessors with improved speech quality.