Bring your own audio

Detector is independent of microphones. It only sees PCM samples, so the audio may come from anywhere.

Required format

Every call to process_audio must contain:

Property	Required value
Sample rate	16,000 Hz
Channels	1 (mono)
Sample type	signed 16-bit PCM (i16)
Block size	160 samples
Time represented	10 ms

The constants SAMPLE_RATE and AUDIO_BLOCK_SAMPLES avoid magic numbers.

use micro_wakeword::{AUDIO_BLOCK_SAMPLES, Detector};

# fn next_block() -> [i16; AUDIO_BLOCK_SAMPLES] { [0; AUDIO_BLOCK_SAMPLES] }
fn main() -> micro_wakeword::Result<()> {
    let mut detector = Detector::from_config("wake-word.json")?;

    loop {
        let block: [i16; AUDIO_BLOCK_SAMPLES] = next_block();
        if let Some(hit) = detector.process_audio(&block)? {
            println!("Detected {}", hit.wake_word);
        }
    }
}

Where app audio comes from

Examples include:

• a WebRTC or voice-call SDK delivering decoded frames;
• a Discord/voice bot receiving Opus packets, then decoding them;
• a browser sending audio over WebSocket;
• an FFmpeg/GStreamer decoding pipeline;
• a game engine’s audio capture callback;
• an embedded or network microphone;
• another Rust capture crate;
• a WAV or raw PCM file.

For compressed data, decode it first. For stereo or another sample rate, downmix and resample before calling the detector.

WebSocket bytes → decode Opus → downmix → resample to 16 kHz → chunks of 160 i16 → Detector

Backpressure and boundaries

Preserve chronological order and do not fabricate continuity after dropped data. If blocks are lost, a call ends, or you switch speakers, call detector.reset() before continuing.

The detector has no cooldown. Low-level callers should use media timestamps or their own state machine if repeated detections need suppression.