Bare-metal audio ML classifier on microcontrollers: case study

A TinyML audio-classification workflow that trains in Python, exports to C and runs directly inside bare-metal firmware.

This DEV.to version is a short engineering note extracted from the case study, with the complete English page linked at the end.

Stack at a glance

Python, C99, scikit-learn, librosa, Random Forest, MFCC, ESP32, STM32.

Why this matters

• Audio ML on microcontrollers is interesting when inference must happen locally, with low latency and no cloud dependency.
• For small classifiers, exporting a trained model into deterministic C can be more practical than embedding a large runtime.

Architecture notes

• Keep the training workflow and firmware integration connected: dataset, feature extraction, model export and C inference should be repeatable.
• Use a compact feature pipeline such as MFCCs only if the memory and timing budget still works on the target MCU.
• Avoid dynamic allocation in the inference path when the firmware has to remain predictable.

Practical takeaways

• TinyML is a system problem. Dataset quality, feature extraction and firmware constraints matter as much as model accuracy.
• Readable generated C is valuable because it can be reviewed, tested and rebuilt when the dataset changes.

Read the full case study

The English page on the Silicon LogiX website contains the full context, visuals and project details: Bare-metal audio ML classifier on microcontrollers.

I keep the DEV.to version intentionally shorter so the canonical page remains the source for the complete case study.