E-Locus - Institutional Repository of the University of Crete

Home Collections School/Department School of Sciences and Engineering Department of Computer Science Post-graduate theses

Post-graduate theses

Current Record: 45 of 824

[Add to Basket]

Identifier

000452055

Title

A neural-based sinusoidal vocoder

Alternative Title

Ένας νευρωνικός ημιτονοειδής φωνοκωδικοποιητής

Author

Ραπτάκης, Μιχαήλ Γ.

Thesis advisor

Στυλιανού, Ιωάννης

Reviewer

Κομοντάκης, Νικόλαος
Πανταζής, Ιωάννης

Abstract

The new era of voice encoding is entirely dominated by neural network models capable of producing natural-sounding synthetic speech, undoubtedly superior compared to all previous parametric methods. However, their exceptionally high quality comes at the cost of being spatially large and computationally demanding. Additionally, despite taking into account some statistical traits of speech signals, most state-of-the-art architectures rarely consider fundamental and well-studied characteristics or methodologies discovered by speech processing literature of the past. In this work, instead of directly synthesizing speech signals by solely using the “raw power” of neural networks, the aim is to take advantage of speech’s quasiperiodic and sinusoidal properties to show how a modern neural-based vocoder can generate speech based on a sinusoidal representation. Using MelGAN as our starting vocoder model due to its renowned speed and quality, we extend it by adding layers that, instead of directly outputting the speech waveform itself, estimate the amplitudes and phases of a new proposed sinusoidal representation. Our results show that the produced quality is on par with the original MelGAN model in terms of MOS scores, indicating that this novel and less expensive approach is indeed feasible. We further experiment with these models and broach the difficulty of finding a multi-resolution spectral loss able to produce quality up to the standards of adversarially-trained models.

Language

English

Subject

Deep learning

Neural networks

Βαθιά μάθηση

Νευρωνικά δίκτυα

Issue date

2022-12-02

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses