Implementation of an AI-based Model for Detecting Speed Variations by Means of Spectrogram Images for New Magnetic Tapes Preservation Strategies - Thesis Summary

Aug 6, 2026

Magnetic tapes used for old audio storage have become obsolete, yet the recorded material remains valuable and is threatened by degradation over time. This study addresses reproduction errors caused by using a playback speed different from the recording speed, which produces inaccurate signals and damages associated cultural heritage. Instead of direct time-domain audio analysis, the work computes spectrograms and applies convolutional neural networks in the frequency domain. Results indicate promising performance and justify further development.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages78

Reading Time3 h 17 min

AccessFree to download

Abstract

Acknowledgements

Introduction

Centro di Sonologia Computazionale

MPAI

Problem and Solution

Irregularities in Playback and Recording Speed

A Possible Solution

Spectrograms

Convolutional Neural Networks

Dataset Creation

Audio Dataset

Spectrogram Dataset

Model Training

Network Configuration

Quality Measures

Comparing Input Sizes

Comparing Scales

Assessment

Testing

Test Dataset

Performance On Test Dataset

Implementation

Future Developments

Conclusion

Appendixes

Code

DEPARTMENT OF INFORMATION ENGINEERING MASTER’S DEGREE IN COMPUTER ENGINEERING IMPLEMENTATION OF ANAI-BASED MODEL FOR DETECTING SPEED VARIATIONS BY MEANS OF SPECTROGRAM IMAGES FOR NEW MAGNETIC TAPES PRESERVATION STRATEGIES Supervisor: Prof. Sergio Canazza Targon co-Supervisors: dr. Alessandro Russo, dr. Matteo Spanio Graduand: Lorenzo Lunardon ACADEMIC YEAR 2023-2024 Abstract Old audio storing methods such as magnetic tapes have been obsolete for decades, now more than ever with the advent of digital formats. The material stored on those tapes, though, is still precious today as it was when it was recorded, and it is at risk of being lost to time, as the materials degrade. The obvious solution to this problem is to digitize these documents, which is a tedious and error-prone process for any technician. This study focuses on preventing one of those errors, which is reproducing the tape using a playback speed different from the recording speed, leading to a very inaccurate signal and a spoilage of the heritage associated to the original document. Instead of analyzing the audio signal directly in the time domain, we opted to explore the frequency domain, computing the spectrograms of the tapes and analyzing them using convolutional neural networks. The approach gave promising results, and the research showed that this is a valid way, worth exploring further. Acknowledgements First and foremost, I would like to thank my family for allowing me to take on this strenuous academic path and being by my side for all bumps in the road. I don’t know if I would have reached the ending of it without their unrelenting support and belief in me, which I myself lacked many times. I also want to express my gratitude to all the friends who have made this whole journey much more pleasurable, from the ones I’ve met in my hometown as a kid and have been with me all along, to the ones I’ve met in Erasmus or some random hostel in the middle of nowhere. I am thankful for the laughs we had together and wish for many more to come. Last but not least, I would like to thank professor Sergio Canazza and Matteo and Alessandro at CSC for helping me write this thesis and letting me contribute to their amazing work. Again, thank you. Contents Abstract I Acknowledgements III 1 Introduction 1 1.1 Centro di Sonologia Computazionale .................. 2 1.2 MPAI ................................... 3 2 Problem and Solution 7 2.1 Irregularities in Playback and Recording Speed ............ 7 2.2 A Possible Solution ............................ 8 2.2.1 Spectrograms ........................... 8 2.2.2 Convolutional Neural Networks ................. 10 3 Dataset Creation 17 3.1 Audio Dataset ............................... 17 3.2 Spectrogram Dataset ........................... 18 4 Model Training 23 4.1 Network Configuration .......................... 23 4.2 Quality Measures ............................. 24 4.3 Comparing Input Sizes .......................... 25 4.4 Comparing Scales ............................. 27 4.4.1 Assessment ............................ 29 5 Testing 33 5.1 Test Dataset ................................ 33 5.2 Performance On Test Dataset ...................... 34 6 Implementation 37 6.1 Future Developments ........................... 39 7 Conclusion 41 Appendixes 45 A Code 45 A.1 Methods used for dividing spectrograms in half ............ 45 A.2 Script used for computing the spectrograms .............. 49 A.3 Main audio analyzer script ........................ 52 A.4 Jupyter Notebook for training the models ............... 56 A.5 Jupyter Notebook for testing the models ................ 63 List of Figures 1.1 Damaged and corrupted magnetic tapes ................ 3 1.2 MPAI AI Framework ........................... 4 1.3 ARP Framework ............................. 4 2.1 Spectrograms of the same sample with three different scales ..... 9 2.2 Mode

FAQs

Why is playback speed mismatch a problem for magnetic tape reproduction?

Reproducing a tape with a playback speed different from the recording speed leads to an inaccurate signal and causes spoilage of the heritage tied to the original document.

What approach does the study use instead of analyzing audio in the time domain?

The study generates spectrograms for the tapes and analyzes them using convolutional neural networks in the frequency domain.

What overall outcome does the research report?

The approach produced promising results, showing that spectrogram-based analysis with CNNs is a valid direction worth further exploration.

Aug 6, 2026

Upload by Lucas Martin