A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms

Filippo Villani; Michele Scarpiniti; Aurelio Uncini

doi:10.1007/978-981-96-0994-9_25

A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms

Filippo Villani, Michele Scarpiniti^*, Aurelio Uncini

^*Kontaktforfatter

Publikation: Bidrag til bog/antologi/rapport/konference proceeding › Konferenceartikel i proceeding › Forskning › peer review

Abstract

In this work, we propose a neural network approach for speech reconstruction from mel spectrograms, a crucial task in achieving high-quality data after processing speech signals in the time-frequency domain. Specifically, we propose a two-stage deep learning approach based on an overcomplete deep autoencoder (DAE) for the mel filter bank inversion coupled with the deep version of the Griffin-Lim (DeGLI) algorithm for the phase information recovery. After the pre-training of both parts of the architecture, a final fine-tuning on the whole system is performed. Some numerical results, evaluated on the well-known TIMIT dataset, demonstrate the effectiveness of the proposed idea by obtaining a PESQ of 3.996, a STOI equal to 0.994, and a mean opinion score evaluated as 4.15.

Originalsprog	Engelsk
Titel	Advanced Neural Artificial Intelligence : Theories and Applications
Redaktører	Anna Esposito, Marcos Faundez-Zanuy, Francesco C. Morabito, Eros Pasero, Gennaro Cordasco
Antal sider	12
Forlag	Springer
Publikationsdato	maj 2025
Sider	267-278
ISBN (Trykt)	978-981-96-0993-2, 978-981-96-0996-3
ISBN (Elektronisk)	978-981-96-0994-9
DOI	https://doi.org/10.1007/978-981-96-0994-9_25
Status	Udgivet - maj 2025
Udgivet eksternt	Ja
Begivenhed	30th International Workshops on Neural Network, WIRN 2023 - Vietri sul Mare, Italien Varighed: 7 jun. 2023 → 9 jun. 2023

Konference

Konference	30th International Workshops on Neural Network, WIRN 2023
Land/Område	Italien
By	Vietri sul Mare
Periode	07/06/2023 → 09/06/2023
Sponsor	International Institute for Advanced Scientific Studies (IIASS)# Department of Psychology, Università della Campania “Luigi Vanvitelli”, IT# Provincia di Salerno# Comune di Vietri sul Mare# International Neural Network Society (INNS)# Università Mediterranea di Reggio Calabria# Società Italiana Reti Neuroniche (SIREN)#

Navn	Smart Innovation, Systems and Technologies
Vol/bind	428
ISSN	2190-3018

Bibliografisk note

Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

Adgang til dokumentet

10.1007/978-981-96-0994-9_25

AUB Link

Søg efter materialet i Aalborg Universitetsbiblioteks søgemaskine

Andre filer og links

Link to publication in Scopus

Citationsformater

Villani, Filippo ; Scarpiniti, Michele ; Uncini, Aurelio. / A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms. Advanced Neural Artificial Intelligence: Theories and Applications. red. / Anna Esposito ; Marcos Faundez-Zanuy ; Francesco C. Morabito ; Eros Pasero ; Gennaro Cordasco. Springer, 2025. s. 267-278 (Smart Innovation, Systems and Technologies, Bind 428).

@inproceedings{af650114d792427481ad1ed6a2f02ce6,

title = "A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms",

abstract = "In this work, we propose a neural network approach for speech reconstruction from mel spectrograms, a crucial task in achieving high-quality data after processing speech signals in the time-frequency domain. Specifically, we propose a two-stage deep learning approach based on an overcomplete deep autoencoder (DAE) for the mel filter bank inversion coupled with the deep version of the Griffin-Lim (DeGLI) algorithm for the phase information recovery. After the pre-training of both parts of the architecture, a final fine-tuning on the whole system is performed. Some numerical results, evaluated on the well-known TIMIT dataset, demonstrate the effectiveness of the proposed idea by obtaining a PESQ of 3.996, a STOI equal to 0.994, and a mean opinion score evaluated as 4.15.",

keywords = "Deep autoencoder, Mel spectrogram inversion, Phase reconstruction, Speech enhancement, Time-frequency representation",

author = "Filippo Villani and Michele Scarpiniti and Aurelio Uncini",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.; 30th International Workshops on Neural Network, WIRN 2023 ; Conference date: 07-06-2023 Through 09-06-2023",

year = "2025",

month = may,

doi = "10.1007/978-981-96-0994-9_25",

language = "English",

isbn = "978-981-96-0993-2",

series = "Smart Innovation, Systems and Technologies",

publisher = "Springer",

pages = "267--278",

editor = "Anna Esposito and Marcos Faundez-Zanuy and Morabito, {Francesco C.} and Eros Pasero and Gennaro Cordasco",

booktitle = "Advanced Neural Artificial Intelligence",

address = "Germany",

}

Villani, F, Scarpiniti, M & Uncini, A 2025, A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms. i A Esposito, M Faundez-Zanuy, FC Morabito, E Pasero & G Cordasco (red), Advanced Neural Artificial Intelligence: Theories and Applications. Springer, Smart Innovation, Systems and Technologies, bind 428, s. 267-278, 30th International Workshops on Neural Network, WIRN 2023, Vietri sul Mare, Italien, 07/06/2023. https://doi.org/10.1007/978-981-96-0994-9_25

A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms. / Villani, Filippo; Scarpiniti, Michele; Uncini, Aurelio.
Advanced Neural Artificial Intelligence: Theories and Applications. red. / Anna Esposito; Marcos Faundez-Zanuy; Francesco C. Morabito; Eros Pasero; Gennaro Cordasco. Springer, 2025. s. 267-278 (Smart Innovation, Systems and Technologies, Bind 428).

Publikation: Bidrag til bog/antologi/rapport/konference proceeding › Konferenceartikel i proceeding › Forskning › peer review

TY - GEN

T1 - A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms

AU - Villani, Filippo

AU - Scarpiniti, Michele

AU - Uncini, Aurelio

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

PY - 2025/5

Y1 - 2025/5

N2 - In this work, we propose a neural network approach for speech reconstruction from mel spectrograms, a crucial task in achieving high-quality data after processing speech signals in the time-frequency domain. Specifically, we propose a two-stage deep learning approach based on an overcomplete deep autoencoder (DAE) for the mel filter bank inversion coupled with the deep version of the Griffin-Lim (DeGLI) algorithm for the phase information recovery. After the pre-training of both parts of the architecture, a final fine-tuning on the whole system is performed. Some numerical results, evaluated on the well-known TIMIT dataset, demonstrate the effectiveness of the proposed idea by obtaining a PESQ of 3.996, a STOI equal to 0.994, and a mean opinion score evaluated as 4.15.

AB - In this work, we propose a neural network approach for speech reconstruction from mel spectrograms, a crucial task in achieving high-quality data after processing speech signals in the time-frequency domain. Specifically, we propose a two-stage deep learning approach based on an overcomplete deep autoencoder (DAE) for the mel filter bank inversion coupled with the deep version of the Griffin-Lim (DeGLI) algorithm for the phase information recovery. After the pre-training of both parts of the architecture, a final fine-tuning on the whole system is performed. Some numerical results, evaluated on the well-known TIMIT dataset, demonstrate the effectiveness of the proposed idea by obtaining a PESQ of 3.996, a STOI equal to 0.994, and a mean opinion score evaluated as 4.15.

KW - Deep autoencoder

KW - Mel spectrogram inversion

KW - Phase reconstruction

KW - Speech enhancement

KW - Time-frequency representation

UR - http://www.scopus.com/inward/record.url?scp=105007745642&partnerID=8YFLogxK

U2 - 10.1007/978-981-96-0994-9_25

DO - 10.1007/978-981-96-0994-9_25

M3 - Article in proceeding

SN - 978-981-96-0993-2

SN - 978-981-96-0996-3

T3 - Smart Innovation, Systems and Technologies

SP - 267

EP - 278

BT - Advanced Neural Artificial Intelligence

A2 - Esposito, Anna

A2 - Faundez-Zanuy, Marcos

A2 - Morabito, Francesco C.

A2 - Pasero, Eros

A2 - Cordasco, Gennaro

PB - Springer

T2 - 30th International Workshops on Neural Network, WIRN 2023

Y2 - 7 June 2023 through 9 June 2023

ER -

Villani F, Scarpiniti M, Uncini A. A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms. I Esposito A, Faundez-Zanuy M, Morabito FC, Pasero E, Cordasco G, red., Advanced Neural Artificial Intelligence: Theories and Applications. Springer. 2025. s. 267-278. (Smart Innovation, Systems and Technologies, Bind 428). doi: 10.1007/978-981-96-0994-9_25

A Two-Stage Neural Network for Speech Signal Reconstruction from Mel Spectrograms

Abstract

Konference

Bibliografisk note

Adgang til dokumentet

AUB Link

Andre filer og links

Fingeraftryk

Citationsformater