TensorFlow Lite Micro on ESP32-S3: A Deployment Case Study in On-Device Audio Event Classification
Year: 2026
Type: article
Abstract: TinyML deployment guides tend to stop at the point where a model converts cleanly to TensorFlow Lite Micro and runs on a microcontroller. Audio classification pipelines rarely stop there, because the feature-extraction stage also has to be ported to embedded C/C++ and kept numerically consistent with the Python training code. This paper documents the deployment of a binary audio event classifier on an ESP32-S3, built around a 57-dimensional MFCC feature vector computed entirely on-device, with no external DSP library. Two inference paths were considered during development. The initial implementation used the standard TensorFlow Lite Micro interpreter with an embedded .tflite model, an operator resolver, and a static tensor arena. This implementation was not completed because its feature-extraction routine remained incomplete and the input-normalization parameters had not been populated. The final firmware instead uses a hand-written C++ forward pass with weights exported directly into a generated header file, eliminating the TensorFlow Lite Micro runtime dependency. The case study highlights two practical considerations for embedded audio machine learning. First, model execution may represent only part of the deployment effort when feature extraction must be independently ported to the target platform. Second, numerical parity between the training and embedded feature pipelines should be checked explicitly rather than inferred from successful compilation or functional device output. The study is limited to one project and does not report accuracy, latency, memory consumption, or flash utilization for the final firmware.
Authors: Nelson R Varte, Diganta Baishya
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7208824729