Generative Models for Improved Naturalness, Intelligibility, and Voicing of Whispered Speech

This work adapts two recent architectures of generative models and evaluates their effectiveness for the conversion of whispered speech to normal speech. We incorporate the normal target speech into the training criterion of vector-quantized variational autoencoders (VQ-VAEs) and Mel-GANs, thereby c...

Full description

Saved in:
Bibliographic Details
Published in:2022 IEEE Spoken Language Technology Workshop (SLT) pp. 943 - 948
Main Authors: Wagner, Dominik, Bayerl, Sebastian P., Maruri, Hector A. Cordourier, Bocklet, Tobias
Format: Conference Proceeding
Language:English
Published: IEEE 09.01.2023
Subjects:
Adaptation models
Cepstral analysis
Conferences
Distortion
Distortion measurement
GAN
generative models
speech conversion
Task analysis
Training
VAE
whispered speech
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This work adapts two recent architectures of generative models and evaluates their effectiveness for the conversion of whispered speech to normal speech. We incorporate the normal target speech into the training criterion of vector-quantized variational autoencoders (VQ-VAEs) and Mel-GANs, thereby conditioning the systems to recover voiced speech from whispered inputs. Objective and subjective quality measures indicate that both VQ-VAEs and MelGANs can be modified to perform the conversion task. We find that the proposed approaches significantly improve the Mel cepstral distortion (MCD) metric by at least 25% relative to a Disco-GAN baseline. Subjective listening tests suggest that the MelGAN-based system significantly improves naturalness, intelligibility, and voicing compared to the whispered input speech. A novel evaluation measure based on differences between latent speech representations also indicates that our MelGAN-based approach yields improvements relative to the baseline.
DOI:10.1109/SLT54892.2023.10022796