← back to work TO BUILD · PLANNED

Edge-AI Audio Analysis

// personal project · edge AI · planned

DOMAINAudio analysis · ML inference

WHEREOn-device / at the edge

TARGETEmbedded (MCU / Zynq — TBD)

STATUSPlanned — to build

The idea

Run audio analysis where the audio actually is. Instead of streaming a sound stream to a server to be classified, this project puts a trained neural model directly on embedded hardware so the device itself can detect or classify what it hears — keeping latency low, working offline, and never sending raw audio off-device.

Why this one

It sits at the intersection of two things I already work in: real-time audio on embedded targets, and deploying ML to constrained hardware. My published work took a quantized model (MobileNetV2, TFLite) onto an STM32 for industrial anomaly detection — this project carries that same edge-AI discipline into the audio domain.

Planned goals

Pick a concrete audio task (e.g. sound-event detection / acoustic anomaly / keyword spotting) and a dataset.
Train and quantize a small model, then convert it for on-device inference (TFLite / TFLite Micro).
Deploy to embedded hardware and measure the real numbers — latency, memory, accuracy after quantization.
Build the audio front end (capture → features → inference) as a clean, real-time pipeline.
Document it as a build-in-public series, the same way as AuraScope.

Scope note: this is the next project on the bench — the specific task, model, and target board will be pinned down as it kicks off. The framing here is the direction, not a finished spec.

Status

Planned. Step-by-step write-ups will appear here as the build starts.