Adaptive Noise Cancellation for Hearing Aids

ISEF Category: Embedded Systems

Subcategory: Signal Processing  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Cafeterias are brutal for hearing aids. Voices overlap, plates clatter, and the background never sits still. That makes speech separation a real engineering problem, not just a volume problem. Your project can test whether a tiny model on a microcontroller can pull speech out of that mess.

What Is It?

This project asks a simple question with a hard answer, can a small device hear the difference between speech and noise well enough to help someone follow a conversation? A hearing aid does not just make sound louder. It has to decide what sound matters, what sound to reduce, and how to do that fast enough to keep speech natural.

Think of it like sorting laundry in the dark. The device gets a mixed pile of sounds, then tries to separate the useful parts from the distractions. A recurrent network is a type of neural network that remembers recent sound patterns, which helps with changing noise. Permutation-invariant loss is a training method that lets the model learn separation even when the order of output streams does not matter. That matters when the model is trying to split speech from babble, since the goal is clean output, not a fixed label order.

Why This Is a Good Topic

This is a strong science fair topic because you can test clear inputs and outputs. You can compare speech clarity, signal-to-noise ratio, latency, and battery-friendly model size across different algorithms or model settings. The topic also connects to a real need, since many people struggle to understand speech in loud places. You can learn signal processing, machine learning, embedded deployment, and experimental design in one project.

Research Questions

• How does a small recurrent network compare with a traditional noise-reduction filter for preserving speech clarity in cafeteria noise?
• What is the effect of model size on speech separation quality and MCU inference speed?
• Does training with permutation-invariant loss improve separation accuracy compared with a standard supervised loss?
• To what extent does the algorithm keep consonants more intelligible than vowels in mixed speech and babble?
• Which noise type, cafeteria babble, classroom chatter, or music, causes the largest drop in separation performance?
• How does quantization from floating-point to integer inference change accuracy, latency, and power use?

Basic Materials

• STM32H7 development board or equivalent MCU board with audio support.
• Microphone module or audio input board.
• Small speaker or headphone output path for playback testing.
• Laptop for model training and data analysis.
• Free audio editor such as Audacity for inspecting recordings.
• Headphones for listening tests.
• SD card or external storage for sample audio files.
• USB cable and board documentation.
• Python installed on a laptop for preprocessing and evaluation.

Advanced Materials

• STM32H7 development board with audio codec and debugging access.
• Development microphone array or calibrated measurement microphone.
• Audio interface for clean recording and playback.
• Oscilloscope or logic analyzer for latency checks.
• Current-measurement tool for power profiling.
• Reference workstation or lab computer for training and export.
• Dataset of speech plus cafeteria or babble noise clips.
• MATLAB or Python environment for signal analysis.
• 3D-printed or foam enclosure for acoustic testing.
• Acoustic meter for checking repeatable test conditions.

Software & Tools

• Python: Preprocess audio, train baseline models, and score separation quality.
• PyTorch: Build the recurrent network and test permutation-invariant training.
• TensorFlow Lite for Microcontrollers: Export a compact model for embedded inference.
• Audacity: Inspect waveforms and compare clean, noisy, and processed clips.
• ImageJ: Measure spectrogram screenshots if you need a simple image-based comparison workflow.

Experiment Steps

1. Define the exact speech problem you will solve, such as separating one voice from cafeteria-like background noise.
2. Choose one baseline method and one embedded neural approach so you can compare a simple filter against the learned model.
3. Build a repeatable audio test set with clear speech, mixed speech, and several noise types.
4. Decide your evaluation metrics, such as intelligibility score, signal-to-noise ratio, inference time, and power draw.
5. Plan how you will reduce the model for the MCU, including quantization and memory limits.
6. Design controls that test whether gains come from real separation, not just louder output or overfitting to one recording set.

Common Pitfalls

• Using only one speaker and one noise sample, which makes the model look better than it really is.
• Testing in a quiet room after training on cafeteria audio, which hides failure in real noise.
• Comparing outputs by ear only, which makes the results hard to defend with data.
• Ignoring latency on the STM32H7, which can make the system useless even if accuracy looks strong.
• Forgetting quantization effects, which can cut accuracy after you move the model from training code to embedded inference.

What Makes This Competitive

A strong project would compare more than one separation strategy and measure more than one outcome. You could test how model compression changes both speech clarity and device speed, then show where the tradeoff breaks down. You could also study which noise types hurt performance most and explain why with signal-level evidence. Careful evaluation, clean controls, and a clear design choice can make the project feel much more like real engineering.

Project Variations

• Test the same separation model on classroom chatter instead of cafeteria babble.
• Compare recurrent networks with a small CNN or classic spectral subtraction method on the same audio set.
• Analyze how fixed-point quantization changes performance across quiet speech, crowded speech, and music background noise.

Learn More

• NIH PubMed: Search review articles on speech separation, hearing aids, and noise suppression to find clinical context and evaluation methods.
• IEEE Xplore: Search for peer-reviewed papers on embedded audio processing and neural speech enhancement if your school has access.
• STM32Cube documentation: Read the MCU vendor guides for audio interfaces, memory limits, and model deployment details.
• MIT OpenCourseWare, Signals and Systems: Review core ideas in filtering, frequency response, and sampling through free course materials.
• TensorFlow Lite for Microcontrollers documentation: Learn how to convert and run small models on embedded hardware.
• NOAA and NASA audio processing resources: Use free signal processing examples and data analysis tools where relevant to waveform work.