Tiny Keyword Spotter on Cortex-M0

ISEF Category: Embedded Systems

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

The Hook

Your phone can hear you because its speech model has room to spare. A Cortex-M0 does not. It has tiny memory, low power, and almost no margin for wasted bits. That makes keyword spotting on a microcontroller a hard and very cool compression problem.

What Is It?

A keyword spotter listens for a small set of words, like a wake word or command word, and ignores everything else. Think of it like a guard at a door. It does not need to understand the full conversation. It only needs to catch one or two key sounds fast enough to react.

This project mixes two ideas. Knowledge distillation means you train a small model using guidance from a much larger speech model. The large model acts like a coach. Mixed-precision activation quantization means you store some model values with fewer bits than others, so the model uses less memory and runs faster on a tiny chip. Your main job is to find the best tradeoff between accuracy, speed, and memory.

Why This Is a Good Topic

This is a strong science fair topic because you can test real engineering tradeoffs, not just build something that works once. You can measure memory footprint, inference time, power use, and recognition accuracy. That gives you clear numbers to compare different model designs. The topic connects to smart speakers, hearing aids, wearables, and other devices that need local speech processing without sending audio to the cloud.

Research Questions

• How does mixed-precision activation quantization affect keyword accuracy on a Cortex-M0?
• What is the effect of knowledge distillation on model size and recall for rare keywords?
• Does per-layer bit-width selection improve accuracy more than a single global quantization scheme?
• To what extent does lowering RAM use change inference latency on the microcontroller?
• Which training setup gives the best balance between false positives and missed detections?
• How does the number of target keywords affect performance under the same memory limit?

Basic Materials

• Cortex-M0 development board with less than 16 kB RAM.
• Microphone module or USB audio input adapter.
• Laptop or desktop computer for training and deployment.
• Headphones or speaker for playback testing.
• Dataset of short speech clips for keywords and background noise.
• MicroSD card or file transfer method for storing test audio.
• Python installed on a computer for model training and analysis.
• Jupyter Notebook for comparing accuracy, latency, and memory results.

Advanced Materials

• Cortex-M0 board with current measurement access.
• Oscilloscope or logic analyzer for timing verification.
• Bench power supply or power monitor for energy testing.
• Audio interface with known sampling characteristics.
• Development kit for cross-compiling embedded code.
• Reference large speech model for teacher training.
• Dataset with labeled wake words and non-keyword speech.
• Profiler or serial logging setup for embedded inference traces.

Software & Tools

• Python: Trains models, runs experiments, and analyzes accuracy and memory tradeoffs.
• PyTorch: Builds the teacher and student speech models for distillation.
• TensorFlow Lite for Microcontrollers: Helps deploy a tiny model on the Cortex-M0 target.
• Jupyter Notebook: Organizes training results, plots confusion matrices, and compares settings.
• ImageJ: Not needed here, so skip it unless you create visual plots from spectrogram images.

Experiment Steps

1. Define the exact keyword task, the memory limit, and the success metrics you will compare.
2. Choose a teacher model, a smaller student model, and a baseline compression method for comparison.
3. Plan how you will represent audio, since feature choice affects both accuracy and microcontroller cost.
4. Design your quantization scheme so different layers can use different bit widths, then decide how you will test each layer choice.
5. Set up controls for noise, speaker variation, and repeated trials so your results reflect real use, not one lucky recording.
6. Build a comparison plan that tracks accuracy, false positives, latency, flash use, and RAM use on the same test set.

Common Pitfalls

• Training only on clean speech, which makes the spotter fail in noisy rooms.
• Choosing a model that fits flash memory but still exceeds RAM during inference.
• Measuring accuracy on the training set, which hides overfitting and gives unrealistically good results.
• Changing audio preprocessing between runs, which makes quantization results impossible to compare.
• Ignoring false positives, which can make a model seem good even when it wakes up on random speech.

What Makes This Competitive

A strong version of this project does more than shrink a model. It shows why your compression method works and where it fails. You could compare your mixed-precision scheme against standard 8-bit quantization, then test the same model across different speakers, noise levels, and keyword sets. Clear error analysis, careful timing data, and a real embedded deployment story make the project much stronger.

Project Variations

• Test the same keyword spotter on wake words for different accents or speaker ages.
• Compare your mixed-precision scheme against post-training 8-bit quantization and training-aware quantization.
• Replace the speech dataset with non-speech audio commands, like alarms or appliance sounds, and measure whether the compression strategy still holds.

Learn More

• TensorFlow Lite for Microcontrollers Guide: Search the TensorFlow documentation for embedded keyword spotting examples and microcontroller deployment notes.
• MIT OpenCourseWare, Deep Learning: Search MIT OpenCourseWare for lectures on neural network compression and model optimization.
• NIH NIDCD Speech and Hearing Resources: Search the NIH National Institute on Deafness and Other Communication Disorders site for speech recognition background and hearing research context.
• PubMed: Search for review articles on speech recognition, quantization, and on-device machine learning.
• IEEE Xplore: Search for papers on keyword spotting, knowledge distillation, and mixed-precision quantization in embedded systems.