🗣️ Speech Enhancement Using Deep Learning

Clean and Noisy Parallel Speech Dataset — Valentini-Botinhao

---

📘 Project Overview

This project aims to enhance noisy speech signals using deep learning models.
It is based on the Valentini-Botinhao Noisy Speech Dataset, which contains parallel pairs of clean and noisy audio files recorded from multiple speakers at 48 kHz.

Due to the large size of the dataset (over 71,000 audio files and 21 GB of data), a subset was extracted for faster experimentation:

• Training set: 100 clean + 100 noisy files
• Test set: 10 clean + 10 noisy files

The main goal is to build, train, and evaluate a speech enhancement model capable of learning to remove background noise from human speech recordings.

---

📊 Dataset Information

Source: Kaggle – Valentini Noisy Speech Dataset

Full Dataset Size:

• 71,000+ files
• ~21 GB total
• Sampling rate: 48 kHz
• Speakers: 28 and 56 (two configurations)

Data Composition:

• clean_trainset_28spk_wav – Clean training speech
• noisy_trainset_28spk_wav – Corresponding noisy training data
• clean_testset_wav – Clean test data
• noisy_testset_wav – Noisy test data

Each noisy file corresponds exactly to a clean version (e.g., p1_1.wav clean ↔ p1_1.wav noisy).

Noises Used:

• Speech-shaped noise
• Babble noise
• Environmental noises from the DEMAND database
• Additional details in:
  • Valentini-Botinhao et al., Interspeech 2016
  • Valentini-Botinhao et al., SSW 2016

---

---

🧠 Model Description

3 model architectures were tested:

1. Simple CNN Autoencoder – Baseline model for denoising.
2. ResNet-based Network – Transfer learning model adapted for audio enhancement.
3. DCCRN (Deep Complex Convolutional Recurrent Network)

· Combines complex-valued convolutions with LSTM layers

· Operates in the time-frequency domain, learning both magnitude and phase

· State-of-the-art for speech enhancement tasks

Both models learn to map a noisy waveform to its corresponding clean waveform using MSE loss.

---

🧩 Requirements

pip install torch torchvision torchaudio
pip install librosa numpy matplotlib tqdm

🚀 How to Run

Prepare the data:

python scripts/prepare_data.py

Train the model:

python scripts/train_model.py

Evaluate the model:

python scripts/evaluate_model.py

THANKYOU !