🧠 OtosakuKWS – On-Device Keyword Spotting (KWS) for iOS

OtosakuKWS is a lightweight, privacy-focused keyword spotting engine for iOS, designed to detect speech commands in real time — entirely on device.

It uses a CRNN CoreML model combined with log-Mel spectrograms for fast, accurate, and low-latency voice command recognition.

🎥 Demo

Watch the model running live on iPhone 13:

🚀 Getting Started

1. Install Feature Extractor

This project depends on the OtosakuFeatureExtractor-iOS Swift package, which extracts log-Mel spectrograms in real time using Accelerate.

It also includes a ready-to-use filterbank archive (filterbank.npy, hann_window.npy).

2. Download Pretrained Model

The CRNN model was trained on the keywords: “go”, “no”, “stop”, “yes”

⬇️ Download model archive

Includes:

• CRNNKeywordSpotter.mlmodelc
• classes.txt

🧪 Validation Metrics

The model was trained on a balanced subset of [Google Speech Commands v2], using strong augmentations and class balancing.

🧩 Integration Example

let kws = try OtosakuKWS(
    modelRootURL: modelURL,
    featureExtractorRootURL: featurizerURL,
    configuration: .init()
)

kws.onKeywordDetected = { keyword, confidence in
    print("Detected: \(keyword) [\(confidence)]")
}

let audioInput = AudioStreamer()

// The `onBuffer` callback receives a chunk of audio sampled at 16kHz, mono (1 channel).
// `AudioStreamer` here is a dummy real-time microphone streamer that simulates live input.
audioInput.onBuffer = { buffer in
    Task {
        await kws.handleAudioBuffer(buffer)
    }
}

📬 Need custom commands?

If you need a custom KWS model for your use case — different keywords, languages, or domain-specific speech — feel free to reach out:

📧 otosaku.dsp@gmail.com

🗝️ Keywords

CoreML, keyword spotting, speech commands, offline voice recognition, privacy-first AI, log-Mel spectrogram, iOS speech processing, CRNN, on-device inference, streaming audio, Swift AI