Choose your API

The library has two layers. Neither is “more accurate”; they feed the same detector. The difference is who owns the audio pipeline.

Your input	Recommended API	What it handles
System microphone + JSON	Listener::from_config	Capture, downmix, resample, detection, cooldown
System microphone + .tflite only	Listener::builder	Same audio work; you supply model settings
Your own 16 kHz PCM + JSON	Detector::from_config	Features, inference, smoothing, threshold
Your own 16 kHz PCM + .tflite only	Detector::builder	Detection; you supply model settings

Listener: own the result, not the audio plumbing

Use it for desktop tools, simple assistants, prototypes, and anything that should listen directly to one input device.

let mut listener = micro_wakeword::Listener::from_config("wake-word.json")?;
while let Some(hit) = listener.next_detection()? {
    println!("{}", hit.wake_word);
}
# Ok::<(), micro_wakeword::Error>(())

Detector: own the audio pipeline

Use it when audio already exists inside your application: a decoded file, WebRTC call, game engine, voice bot, browser stream forwarded over a socket, or another capture library.

use micro_wakeword::{AUDIO_BLOCK_SAMPLES, Detector};

# fn run(block: [i16; AUDIO_BLOCK_SAMPLES]) -> micro_wakeword::Result<()> {
let mut detector = Detector::from_config("wake-word.json")?;
if let Some(hit) = detector.process_audio(&block)? {
    println!("{}", hit.wake_word);
}
# Ok(()) }

The low-level API expects exactly 160 signed 16-bit mono samples at 16 kHz per call. It intentionally does not resample, downmix, capture, buffer, or apply a cooldown.

Mix the layers

You can construct a detector yourself and hand it to a listener:

use micro_wakeword::{Detector, Listener};

# fn run() -> micro_wakeword::Result<()> {
let detector = Detector::from_config("wake-word.json")?;
let listener = Listener::from_detector(detector, Some("USB Mic"))?;
# Ok(()) }

This is useful when detector construction needs custom logic but microphone capture should remain automatic.