Machine Learning Noise Cancellation System

Final Year Project submitted for the degree of B.Sc. (Hons.) in Computer Engineering at the UOM under the supervision of Dr. Trevor Spiteri.

---

Project Overview

This project explores the application of machine learning (ML) techniques for speech enhancement, aiming to justify their superiority over classical methods in real-time denoising scenarios.

A modular and extensible Python pipeline was built to support:

• Variable-length waveform inputs using dynamic bucketing
• Real-time inference via spectrogram-based denoising
• Consistent metric-based evaluation for batch and single-file use
• Reproducible experiments via centralised configuration and logging

Three classical denoising methods — Spectral Subtraction (SS), Wiener Filtering (WF), and MMSE-LSA — were implemented to provide baselines.

These were compared against five ML models trained from scratch:

• CNN: A shallow baseline to verify the pipeline
• CED / R-CED: Encoder-decoder variants with and without residuals
• UNet: A deeper skip-connected model for better feature preservation
• Conv-TasNet: A time-domain network yielding top performance

All models were trained on magnitude spectrograms, with batch handling strategies (Static, Dynamic Bucketing, PTO) evaluated separately.

The pipeline supports training from scratch, memory-efficient inference, and consistent evaluation. Laying the foundation for future experimentation with advanced architectures.

---

Evaluation Metrics

• SNR – Signal-to-noise ratio
• MSE – Mean squared error
• PESQ – Perceptual evaluation of speech quality
• STOI – Short-time objective intelligibility
• LSD – Log-spectral distance

The metrics were selected based on relevance in related literature and provide a balanced view of both numerical accuracy and perceptual quality in denoising performance.

---

Model & Method Performance

Method	↑SNR (dB)	↓MSE	↑PESQ	↑STOI	↓LSD (dB)	Time (s)
Baseline	-2.28	0.005152	1.8451	0.8928	0.9042	56
SS	3.09	0.001525	1.4535	0.8457	0.7671	61
WF	0.46	0.002875	2.0639	0.8889	0.7535	73
MMSE-LSA	-0.86	0.003726	2.0238	0.8943	0.7971	83
CNN	4.64	0.001344	1.7410	0.8073	0.7956	71
CED	13.19	0.000161	1.6780	0.8386	0.7655	65
R-CED	14.53	0.000117	2.0542	0.8677	0.6480	74
UNet	16.99	0.000069	2.1384	0.8940	0.7076	87
Conv-TasNet	18.06	0.000063	2.4329	0.9112	0.6741	139

---

Key Findings

• Dynamic Bucketing was selected as the preferred dataset handling method. It used K-Means clustering to assign samples into optimally sized buckets, improving training efficiency over Static Bucketing while avoiding the runtime penalties of Padding-Truncation Output-Truncation (PTO) during inference.

• OOM mitigation techniques — including mixed-precision (FP16), garbage collection, and gradient accumulation — enabled training of deeper models on the university GPU cluster. Evaluation showed no degradation in model quality, and in some cases, slight improvements due to accumulated gradient stability.

• Among classical methods:

  • Spectral Subtraction (SS) achieved strong numerical performance (e.g., SNR), but performed poorly on perceptual metrics like PESQ and STOI.
  • Wiener Filtering (WF) and MMSE-LSA achieved better perceptual quality, but failed to match ML models in numerical fidelity.
  • Overall, no classical approach provided a comprehensive improvement across all evaluation dimensions.

• Conv-TasNet emerged as the top-performing ML model, achieving the best SNR, PESQ, and STOI. Originally developed for speech separation, its temporal masking architecture and learned bottlenecks translated effectively to the spectrogram-based denoising task used in this pipeline.

• The pipeline is built for future extensibility and supports:

  • Integration of transformer and diffusion-based models
  • Real-time inference with beamforming support
  • Exploration of unseen noise generalisation and more diverse datasets

---

Limitations & Future Work

Despite achieving strong numerical and perceptual performance, this project leaves several avenues for improvement:

Limitations

• Perceptual Ceiling: While Conv-TasNet outperformed classical models, its PESQ score (2.43) remains well below the perceptual upper bound of 4.5.
• Generalisation to Unseen Noise: The models struggled with noise types not present during training. More diverse datasets are needed to improve real-world robustness.
• Resource Constraints: Due to limited GPU memory, batch sizes and model complexity were capped. Techniques like gradient accumulation and FP16 were essential but not ideal.

Future Directions

• Model Expansion: Incorporating transformer-based models (e.g. ScaleFormer) or diffusion-based architectures could unlock further gains in intelligibility and naturalness.
• Multi-Channel Input: Extend the pipeline to support beamforming and microphone array input for spatial filtering in real-world deployments.
• Self-Supervised Pretraining: Introduce SSL or reinforcement learning strategies to improve generalisation with limited labelled data.
• Real-Time Integration: Adapt the inference system for on-device deployment in edge hardware like headphones or smartphones with Active Noise Cancellation (ANC) support.

---

Repository Structure

.Project/
├── main.py              # Entry point for training/evaluation
├── config.py            # Central config for datasets, models, training params
├── Utils/
│   ├── dataset.py       # Spectrogram conversion + augmentation
│   ├── denoise.py       # Inference utilities
│   ├── train.py         # Model training/validation logic
│   └── models.py        # Model architectures (CNN, CED, R-CED, UNet, Conv-TasNet)
├── Models/              # Saved model weights by experiment
│   ├── 25/
│   ├── dataset/
│   └── oom/
├── Output/              # Denoising outputs (wav, txt, png)
│   ├── 25/
│   ├── dataset/
│   ├── oom/
│   ├── png/
│   ├── txt/
│   └── wav/
├── Cache/               # Cached spectrograms and length logs
│   ├── dynamic/
│   ├── static/
│   └── pto/
├── ssh/                 # SLURM-compatible job scripts
│   ├── main.sh
│   ├── latex.sh
│   └── notebook.sh
├── Template/            # Report LaTeX source
│   ├── main.pdf         # Final Year Project Report
│   ├── main.tex
│   ├── build/
│   ├── content/
│   └── references.bib
└── .gitignore

---

Dataset

The system uses the Noisy Speech Database from the University of Edinburgh:

• https://datashare.ed.ac.uk/handle/10283/2791
• License: Creative Commons Attribution 4.0 International

Audio is converted to magnitude spectrograms for all training, validation, and inference steps.

---

Usage

All functionality is controlled via main.py and config.py:

python main.py

To modify:

• Dataset or output locations
• Model selection or loss functions
• Batch sizes, precision, or training strategy

Edit config.py accordingly.

Inference can denoise a single .wav or full batch denoice with metric evaluation as a .txt.

---

Final Report

Read the full dissertation here: main.pdf

Includes methodology, system design, ablation studies, and model evaluation.

This project received an A Grade in the B.Sc. (Hons.) Computer Engineering programme.

---

Author

Graham Pellegrini
B.Sc. (Hons.) Computer Engineering
University of Malta
GitHub: @GrahamPellegrini