Go Alarm Detection for Microprocessors

Released a rule-based audio alert detector built in Go.

Target signals include the following.

• police car sirens
• air sirens
• ambulance sirens
• emergency signals
• horns
• house alarms
• smoke detector alarms

Instead of a trained ML model, the rule-based approach was chosen with the goal of running smoothly even on microprocessors.

About 200 sound files were collected from Freesound, Pixabay Sounds, YouTube, and other sources, cut and edited with Audacity, and then an optimization script automatically computes the best-fitting ruleset for these samples. The approach tunes the rules to clearly separate positive and negative samples.

To reduce overfitting, repeated testing was done with audio not included in training, and data augmentation was also applied.

• overlaying ambience sounds at low volume
• adjusting playback speed
• adjusting volume
• automated trimming

The resulting data grew to the level of thousands of samples. It's stated that some LLM was used to write metric calculation code for code generation.

Repo: github.com/therealPaulPlay/go-alert-detector