Model JSON

The JSON file is the safest way to distribute a model because it carries the settings chosen by its author.

{
  "type": "micro",
  "wake_word": "miku",
  "author": "Model author",
  "website": "https://example.com/model",
  "model": "miku.tflite",
  "trained_languages": ["en"],
  "version": 2,
  "micro": {
    "probability_cutoff": 0.3,
    "sliding_window_size": 3,
    "feature_step_size": 10
  }
}

Fields the crate uses

Field	Meaning
type	Must be "micro"
wake_word	Human-readable label returned in a detection
model	.tflite path, relative to this JSON unless absolute
version	Must currently be 2
micro.probability_cutoff	Minimum smoothed score needed to detect
micro.sliding_window_size	Number of recent scores averaged
micro.feature_step_size	Must be 10 milliseconds
author, website, trained_languages	Exposed as model metadata

Extra JSON fields used by ecosystems such as ESPHome are accepted and ignored.

Load it directly

use micro_wakeword::{Config, Detector, Listener};

# fn run() -> micro_wakeword::Result<()> {
let config = Config::from_file("models/miku.json")?;
println!("Model: {}", config.model_path.display());

let detector = Detector::from_config("models/miku.json")?;
let listener = Listener::from_config("models/miku.json")?;
# Ok(()) }

Config::from_file validates values before the model is loaded. Relative model paths are resolved against the JSON file’s directory, so launching the program from another working directory does not break that relationship.