Revolutionizing Speech Enhancement with CBAR

Cleaner audio for Nigeria’s urban noise, powered by AI.

The Dataset is made of over 1300 hours of sound recording CBAR outperformed SEGAN by 6% on Nigerian Urban Noises Noise levels in Nigerian cities often exceed WHO guidelines Recordings from Abuja, Lagos, Port-Harcourt and Uyo Port-Harcourt is a leading center for Noise pollution Designed for non-stationary noise environments Urban Nigerians at 60% risk of harmful noise exposure Ibadan milling workshops: 95–105 dB Akure markets: noise > WHO limit, lower noise = higher sales Ilorin industrial zones up to 110 dB Port Harcourt junctions hit ~90 dB daily noise

Why Urban Clarity Matters

Built for our real-world soundmotsphere; markets, highways, generators and city life.

Everyday Noise

Lagos, Abuja, PHC and Uyo present complex noise patterns that degrade speech clarity.

Health Impact

Noise above 85 dB causes stress and hearing loss. Nigerian traffic often exceeds 90 dB daily.

Economic Loss

Studies show markets with lower noise see higher sales. CBAR helps keep voices clear.

Built for Nigerians

Dataset mirrors our environment for better generalization in real scenarios.

Proven Results

Consistent gains over legacy models on internal benchmarks.

The Dataset

CBAR is trained on a uniquely sourced Nigerian Urban Noise dataset of over 1300 hours of real-world recordings including generator hums, busy market chatter, bus engines and honks, street sounds and urban crowd noise.

This was ethically collected and the voices where recorded, remain anonymous from urban centers such as Port-Harcourt, Abuja, Lagos and Uyo. As such, this noise exists as Non-Stationary and is the primary motivation for the Pre-Processing and Architecture choice.

Data Pre-Processing involved use of Short Time Fourier Transform (STFT) to obtain frequency magnitudes (training data) and Inverse Short Time Fourier Transform (iSTFT) to reconstruct the denoised magnitude and hence, the denoised audio.

Structured Data

1300+ hours grouped into Generators, Traffic & Vehicles, Streets & Markets, plus a combined mega set.

Cities Covered

Port Harcourt
Uyo
Abuja
Lagos

Preprocessing

Short-Time Fourier Transform (STFT) for magnitudes; iSTFT for waveform reconstruction.

How CBAR Works

You might now be wondering, how really does this Denoiser work? Let's break it down and understand the components:

1. Convolutional Neural Networks (CNN)

These are Neural Networks primarily built for work with images but their feature extraction capability serve a core purpose here. By sliding over the time axis of the Magnitude input, it extracts low level features from the input.

2. Bidirectional Long Short-Term Memory (BiLSTM)

LSTM is a Neural Network type gotten from the concept of Recurrent Neural Networks (RNN) which work with sequential time-dependent data. This BiLSTM component allows forward and backward context understanding by the Model.

3. Attention Mechanism

This mechanism enables the model to assign diverse weights to various parts of the input sequence based on their relevance when producing an output. This capability is particularly critical for handling inputs where the lengths and relevance strengths of different parts can vary significantly.

4. Residual Connections

To solve a classic Machine Learning problem of "Vanishing Gradient", Residual Connections have been employed. These improve optimization by bringing forward the input of the layer and adding to the output of that layer hence preventing diminshed signals as they pass through several layers.

Just as the name should imply, these components are sequential coming one after the other. This can be visualized in the model architecture below.

Try CBAR

Experience how CBAR denoises speech. Live API and uploads coming soon.

API Preview

POST /denoise
{ "audio_file": "noisy_sample.wav" }

Response:
{ "status": "success", "clean_audio": "link/to/output.wav" }

Live API integration in progress

Upload a Sample

Soon, you’ll be able to upload your own audio and test CBAR online.

When the Denoiser is finally LIVE, you should expect a denoised audio as well as visualizations such as waveform and spectrogram. Check back soon!

Documentation & Research

Explore publications, blogs and technical papers built on CBAR.

Blog Research Papers Books API

Inside the CBAR Dataset

How we captured 1300+ hours across four cities and prepared the data as well as updates on datasets, methods and real-world use cases.

Conference Paper

Peer-reviewed IEEE-format research paper presenting CBAR as a Noise Reduction approach at the ICMEAS 2025.

View Papers →

The CBAR Handbook

Expanded report documenting CBAR’s journey from background study, literature review up to data collection and deployment.

Explore Books →