Speech Noise Suppression with Deep Learning in MATLAB

by Eric Vo, Kailash Rao, Siwoo Chung, Gabriel Ramos

Overview

This project is a submission for the MATLAB and Simulink Challenge Project: Speech Background Noise Suppression with Deep Learning. Our project fine-tunes and analyzes a deep learning model for speech noise suppression using MATLAB's Audio Toolbox and Deep Learning Toolbox. The goal was to develop a denoising system that improves speech quality in noisy environments and to evaluate its effectiveness using both:

• Subjective methods – listening-based, human-rated clarity
• Objective metrics – quantitative performance (e.g., noise reduction ratio)

Folder Structure

data/
├── noise/                             # Background noise samples used for mixing
├── train/                             # Training dataset
│   ├── clean_trainset_wav/            # Clean training audio
│   └── noisy_trainset_wav/            # Noisy training audio
├── validate/                          # Validation dataset (for future use)
│   ├── clean_validationset_wav/       # Clean validation audio
│   └── noisy_validationset_wav/       # Noisy validation audio
├── test/
│   ├── simpleTest/                    # Small sample test case
│   │   ├── noisyInput/                # Input audio with noise
│   │   └── testOutput/                # Denoised output from simple test
│   ├── gabrielSamples/                # Gabriel's (custom) evaluation samples
│   │   ├── clean/                     # Original clean speech
│   │   ├── clean8Khz/                 # Clean speech resampled (8 kHz)
│   │   ├── noisy8Khz/                 # Noisy input resampled (8 kHz)
│   │   └── output_wav/                # Denoised output files
│   └── datasetSamples/                # General test set used for metrics
│       ├── clean_testset_wav/         # Clean test audio
│       ├── noisy_testset_wav/         # Noisy test audio
│       ├── clean_testset_wav_8khz/    # Clean test audio resampled (8 kHz)
│       ├── noisy_testset_wav_8khz/    # Noisy test audio resampled (8 kHz)
│       └── output_wav/                # Denoised output files

models/
├── denoiseNet_FineTuned_VBD.mat       # Final fine-tuned model
└── denoiseNetFullyConnected.mat       # Original pre-trained model

scripts/
├── main.m                             # Top-level script (demonstrates functions)
├── train.m                            # Script to fine-tune the model
├── generateNoisyFile.m                # Add noise to a single file
├── generateNoisyDir.m                 # Add noise to a directory of files
├── denoiseSpeechFile.m                # Denoise a single file using model
├── denoiseSpeechDir.m                 # Denoise all files in a directory
├── calculateAudioError.m              # Computes RMSE, SNR, PSNR, etc.
├── calculateNoiseReductionRatio.m     # Computes noise suppression ratio
├── resampleDir.m                      # Resamples files for subjective comparison

media/                                 # Images used in README.md

LICENSE                                # Usage restriction notice
README.md                              # Project documentation

Running the Project Example

Important: All functions and scripts only work with .wav audio files.

Running main.m will:

• Fine-tune the pre-trained model
• Add noise to custom audio samples
• Denoise multiple samples using the fine-tuned model
• Calculate the noise reduction ratio between the noisy and denoised audio

How to Run main.m:

1. Add scripts/ to your MATLAB path.
2. (Optional) To enable initial training of the model, uncomment line 7 in main.m (may be resource-intensive).
3. Run the script. The denoised files will be outputted in data/test/gabrielSamples/output_wav/ and data/test/datasetSamples/output_wav/.

Functions

Used in main.m

• train(model, cleanDir, noisyDir, outputPath, options: useGPU=0)
  • Fine-tunes a denoising neural network using paired clean and noisy audio
• generateNoisyDir(cleanInputDir, noiseDir, noisedOutputDir, options: snr_dB=10, clear=1)
  • Generates 5 noisy audio files for every clean audio file
• denoiseSpeechDir(model, noisyInputDir, denoisedOutputDir, clear=1)
  • Denoises a specified directory of noisy audio files
• calculateNoiseReductionRatio(denoisedAudioArray, cleanOriginalDir, noisyOriginalDir, numDenoised=1)
  • Computes noise reduction ratios of denoised over noisy audio

Helper Functions

• generateNoisyFile(cleanPath, noisePath, outputPath, snr_dB)
  • Adds background noise to a clean audio file at a specified SNR
• denoiseSpeechFile(model, noisyInput, denoisedOutputDir)
  • Denoises a single noisy file
• calculateAudioError(cleanAudio, denoisedAudio)
  • Calculates several error metrics (RMSE, MSE, SNR, NRMSE, PSNR, MAE, correlation) by comparing the model's denoised audio with the original clean audio
• resampleDir(inputDir, outputDir, targetFs)
  • Resamples the specified directory of audio files to the desired target frequency (used for subjective evaluation)

See function docstrings for more details

Model & Features

• Model: Pre-trained model from MATLAB's Denoise Speech Using Deep Learning Networks Example
• Training Data: Used the VoiceBank-Demand (VBD) Dataset to fine-tune pre-trained model
• Preprocessing: STFT with Hamming windows
• Feature Context: 8-segment windows
• Sampling Rate: 8 kHz

Metrics from fine-tuning the pre-trained model. The two graphs display the root mean square error (RMSE) and training loss of the model over time.

Dependencies

• MATLAB R2025a or later
• Audio Toolbox
• Deep Learning Toolbox
• Signal Processing Toolbox

Project Timeline

Time Frame: 6/27/2025 – 7/25/2025 (4 weeks)

• Conducted a literature review on deep learning-based speech noise suppression
• Downloaded and prepared speech/noise datasets from the Denoise Speech Using Deep Learning Networks Example
• Prepared our own clean sample data (gabrielSamples)
• Generated noisy samples using custom noise at various SNR levels
• Fine-tuned a pre-trained model on VoiceBank-Demand samples
• Applied STFT preprocessing and neural network inference
• Evaluated denoising performance via:
  • Subjective listening tests (clarity scale: 1–5)
  • Objective metrics (Noise reduction ratio, RMSE, SNR, correlation, PSNR, etc.)

Future Improvements

Given more time and more capable hardware, we would implement the following improvements:

• Develop and train our own Deep Learning Model from scratch using larger datasets
• Use a validation set in conjunction with the training set while fine-tuning model

Results

Objective Results

• Average Noise Reduction Ratio: 0.625 for data/test/gabrielSamples, 0.732 for data/test/datasetSamples

  • The average reduction ratio specifies the proportion of noise the model removed from the noisy files. Positive values indicate less noisy output, while negative values indicate noisier output.
  • On average, 62.5% and 73.2% of the noise was removed for data/test/gabrielSamples and data/test/datasetSamples, respectively.

• Individual Noise Reduction Ratios for data/test/gabrielSamples:

Subjective Ratings (from data/test/gabrielSamples/output_wav/):

File	Clarity (1–5)
10_waves_5dB_dn.wav	3.7
10_pencils_5dB_dn.wav	3.2
10_jet_city_birds_5db_dn.wav	4.0
10_cafe_5db_dn.wav	3.1
10_birds_farm_5db_dn.wav	3.1

Contributors

Name	Contributions	Files Worked On
Siwoo Chung	Noising algorithm, noising test dataset function documentation	main.m, generateNoisyFile.m, generateNoisyDir.m
Kailash Rao	Model tuning, denoising test dataset, aggregating functions in main.m	main.m, train.m, denoiseSpeechFile.m, denoiseSpeechDir.m
Eric Vo	Noise reduction ratio metrics, refactoring functions	main.m, resampleDir.m, calculateNoiseReductionRatio.m, calculateAudioError.m
Gabriel Ramos	Dataset preparation, README documentation, updates on key requirements and deadlines	README.md