Chaotic Reservoir Speech Classifier

ISEF Category: Embedded Systems

Subcategory: Circuits  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A tiny circuit can act like a memory with chaos inside it. That sounds wrong, but it is the whole trick behind reservoir computing. You can turn a spoken digit into a stream of voltage samples, then ask a simple readout to guess what was said.

What Is It?

This project uses a Chua's circuit, a famous chaotic circuit that produces rich, irregular signals, as a physical reservoir. Think of the reservoir like a bowl of bouncing balls. You tap it with an input, and the messy motion spreads that input into many useful patterns. Instead of training the whole circuit, you train only a small linear readout, which is the part that maps the circuit's output to a label, like digit zero through nine.

Your MCU, or microcontroller unit, samples the analog signal with its ADC, which stands for analog-to-digital converter. The ADC turns voltage into numbers. Those numbers become features for classification. The digital side of the comparison uses a small LSTM, a recurrent neural network built to handle sequences. You compare the analog reservoir and the digital model at the same rough size, so you can ask which one gives better accuracy, lower power, or faster inference for spoken digit recognition.

Why This Is a Good Topic

This topic works well because it has a clear input, a clear output, and many knobs you can test. You can change circuit parameters, sampling rate, feature window size, readout method, or noise level, then measure how classification changes. It connects to edge AI, low-power sensing, and hardware that does useful work without a big computer. You can also learn signal processing, machine learning basics, and experimental design in one project.

Research Questions

• How does changing the Chua circuit parameter set affect spoken digit classification accuracy?
• What is the effect of ADC sampling rate on reservoir feature quality and final classification performance?
• Does a linear readout outperform a small ridge regression model on the same reservoir features?
• To what extent does input scaling change the separability of digit classes in the reservoir state space?
• Which window length gives the best tradeoff between accuracy and latency for spoken digit recognition?
• How does the physical reservoir compare with a same-size digital LSTM in accuracy, inference time, and estimated energy use?

Basic Materials

• Microcontroller board with ADC, such as an Arduino Portenta, ESP32, or similar.
• Breadboard or perfboard for circuit prototyping.
• Chua circuit components, including op-amps, resistors, capacitors, and a nonlinear element.
• Oscilloscope or USB logic analyzer for checking the analog waveform.
• Microphone or recorded speech dataset of spoken digits.
• Laptop for data collection, training, and analysis.
• Jumper wires, power supply, and multimeter.
• External ADC module if the MCU ADC resolution is too low.
• Serial cable or programming cable for the MCU.

Advanced Materials

• Bench power supply with current limit.
• Oscilloscope with data export.
• Function generator for controlled input tests.
• Precision resistors and capacitors for tuning the Chua circuit.
• High-speed external ADC evaluation board.
• Low-noise instrumentation amplifier, if needed for signal conditioning.
• FPGA or higher-end MCU for timing comparison.
• Microphone array or calibrated audio playback setup for repeatable speech input.
• Computer with Python, PyTorch, or TensorFlow for baseline model training.

Software & Tools

• Python: Cleans waveform data, extracts features, and trains the baseline classifier.
• Jupyter Notebook: Lets you document trials, plots, and model comparisons in one place.
• NumPy: Handles arrays of ADC samples and reservoir states.
• SciPy: Supports signal filtering, resampling, and statistical tests.
• ImageJ: Can help if you document oscilloscope screenshots and need simple measurement from images.

Experiment Steps

1. Define the exact speech task, such as classifying isolated spoken digits from a fixed dataset or your own recordings.
2. Choose the physical features you will feed into the readout, such as raw ADC samples, summary statistics, or delayed samples.
3. Build a fair baseline by matching the digital LSTM to the same input set, similar parameter budget, and same train-test split.
4. Decide which circuit settings, sampling choices, and input scaling values you will sweep first.
5. Plan a measurement pipeline that captures repeatable reservoir states, labels them correctly, and stores them in a clean format.
6. Set your evaluation metrics before testing, including accuracy, confusion matrix, latency, and estimated energy per prediction.

Common Pitfalls

• Letting the Chua circuit drift between runs, which changes the reservoir state distribution and breaks repeatability.
• Sampling the analog waveform too slowly, which throws away the dynamics that make the reservoir useful.
• Training and testing on speech recordings from the same speaker, which inflates accuracy and hides weak generalization.
• Comparing the analog system to a much larger or much smaller LSTM, which makes the benchmark unfair.
• Feeding noisy or clipped ADC values into the readout, which can make the classifier look unstable for the wrong reason.

What Makes This Competitive

A strong version of this project does more than report one accuracy number. You can test multiple circuit regimes, show why one regime works better, and compare more than one baseline metric. You can also look at energy use, latency, and class confusions, not just overall accuracy. A deeper project explains how the chaotic dynamics help separate speech patterns, then backs that claim with clean data and fair comparisons.

Project Variations

• Use environmental sound clips instead of spoken digits, then test whether the reservoir still separates short acoustic events well.
• Replace the Chua circuit with another nonlinear analog front end and compare which one gives cleaner features for the same classifier.
• Keep the circuit fixed, then study how speaker accent, noise level, or microphone distance changes classification performance.

Learn More

• MIT OpenCourseWare, Signals and Systems: Search MIT OpenCourseWare for courses on sampling, filtering, and time series signals.
• U.S. National Institute of Standards and Technology, digital signal processing resources: Search the NIST site for intro material on ADCs, sampling, and measurement.
• Chaos and Time-Series Analysis: Look for university library access or publisher previews to learn how chaotic signals are analyzed.
• PubMed: Search for review articles on reservoir computing and speech recognition to see how the method is used in research.
• IEEE Xplore: Search for papers on physical reservoir computing, Chua circuits, and edge speech classification through your school or public library access.