micro-wakeword

Local wake-word detection for Rust, powered by microWakeWord models.

micro-wakeword runs microWakeWord-compatible TensorFlow Lite models entirely on-device. Use the ready-made microphone listener, or feed PCM from your own audio pipeline.

• Local inference: no cloud service or network connection
• Simple microphone API with resampling and channel conversion
• Low-level streaming API for files, sockets, bots, and custom audio engines
• Standard microWakeWord JSON support, plus a model-only builder
• Configurable cooldown, threshold, device, and TensorFlow Lite runtime

Microphone ──> Listener ─┐
                        ├──> Detector ──> Detection { wake_word, probability }
Your 16 kHz PCM ─────────┘

[!NOTE] Windows x86-64 is supported out of the box. Other platforms currently need a compatible TensorFlow Lite C shared library supplied by the application.

Install

cargo add micro-wakeword

Windows builds require the MSVC C++ build tools because the audio frontend is compiled as part of the crate.

Quick start: listen to a microphone

use micro_wakeword::Listener;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut listener = Listener::from_config("wake-word.json")?;
    println!("Listening for {}", listener.detector().config().wake_word);

    while let Some(detection) = listener.next_detection()? {
        println!(
            "Detected {} ({:.1}%)",
            detection.wake_word,
            detection.probability * 100.0
        );
    }
    Ok(())
}

The JSON file supplies the model path and its recommended detection settings. Relative model paths are resolved from the JSON file's directory.

Choose a microphone and cooldown

use std::time::Duration;
use micro_wakeword::Listener;

# fn run() -> micro_wakeword::Result<()> {
let mut listener = Listener::config_builder("wake-word.json")?
    .device("USB Mic") // A device index or part of its name.
    .cooldown(Duration::from_millis(500))
    .build()?;
# Ok(()) }

The default cooldown is one second. Use Duration::ZERO to disable repeat suppression. List available inputs with available_input_devices().

Only have a .tflite model?

JSON is recommended, but it is not required. Provide the settings yourself:

use micro_wakeword::Listener;

# fn run() -> micro_wakeword::Result<()> {
let mut listener = Listener::builder("hey_computer.tflite")
    .wake_word("hey computer")
    .probability_cutoff(0.5)
    .sliding_window_size(3)
    .build()?;
# Ok(()) }

The crate cannot infer the cutoff or sliding-window size from a TFLite file. Use values supplied by the model author when possible. Otherwise, validate them against recordings containing the wake word and recordings containing ordinary speech and background noise.

Bring your own audio

Use Detector when audio already comes from a file, WebSocket, voice bot, media pipeline, or another capture library:

use micro_wakeword::{AUDIO_BLOCK_SAMPLES, Detector};

# fn run() -> micro_wakeword::Result<()> {
let mut detector = Detector::from_config("wake-word.json")?;
let pcm = [0_i16; AUDIO_BLOCK_SAMPLES];

if let Some(detection) = detector.process_audio(&pcm)? {
    println!("Detected {}", detection.wake_word);
}
# Ok(()) }

Each call requires exactly 160 samples of 16 kHz, mono, signed 16-bit PCM—10 milliseconds of audio. The low-level detector does not resample, downmix, or apply a cooldown. Call reset() before a new unrelated stream or after an audio discontinuity.

Which API should I use?

Runnable examples

Commands below use the sample files in this repository's ../models folder:

detect_pcm expects headerless little-endian i16 audio at 16 kHz mono. Omit audio.raw to run it against generated silence.

TensorFlow Lite runtime

On Windows x86-64, Runtime::Auto uses this order:

1. MICRO_WAKEWORD_TFLITE_LIB
2. TFLITE_C_LIB
3. The bundled, checksum-verified TensorFlow Lite 2.17.1 runtime

To choose a library explicitly:

use micro_wakeword::{Detector, Runtime};

# fn run() -> micro_wakeword::Result<()> {
let detector = Detector::from_config_with_runtime(
    "wake-word.json",
    Runtime::from_path("tensorflowlite_c.dll"),
)?;
# Ok(()) }

Runtime::System searches the platform locations supported by tflite-c-rs.

Model compatibility

Models must use the standard microWakeWord pipeline:

• 16 kHz mono input audio
• 40 frontend features
• Signed int8 input tensor shaped [1, rows, 40]
• One int8, uint8, or float32 probability output
• Standard configuration format version 2 when using JSON
• 10 ms feature step

An incompatible model returns a descriptive error while loading.

Feature flags

The default listener feature enables CPAL microphone capture and Rubato resampling. Applications supplying their own 16 kHz PCM can leave them out:

micro-wakeword = { version = "0.1", default-features = false }

License

micro-wakeword is available under the MIT License. Bundled TensorFlow microfrontend and KissFFT sources retain their upstream licenses; see LICENSES.