OtosakuStreamingASR is a lightweight on-device streaming speech recognition engine for iOS. It performs real-time audio processing using a Conformer-based architecture and CTC decoding.

• ✅ Fully offline
• 🎙 Real-time streaming speech recognition
• 🛠 Modular architecture (feature extractor, encoder, decoder)

Watch the model running live on iPhone 13:

Add the Swift Package to your Xcode project:

https://github.com/Otosaku/OtosakuStreamingASR-iOS

import OtosakuStreamingASR
                                                                                                
let asr = OtosakuStreamingASR()
try asr.prepareModel(from: modelURL)
asr.subscribe { text in
    print("🗣 Recognized: \(text)")
}
// Raw audio chunk: [Double] in range [-1.0, 1.0], strictly 2559 samples per chunk (80ms at 16kHz)
try asr.predictChunk(rawChunk: yourRawAudioChunk)
try asr.stop() // Finalize and decode remaining buffer
asr.reset() // Reset internal model state

• Architecture: Fast Conformer (Cache-Aware Streaming)

• Language: 🇷🇺 Russian (fine-tuned from English)

• Training: 250 hours of Russian speech (30 epochs)

• WER (Word Error Rate):

  • Russian (fine-tuned): 11%
  • English (before fine-tuning): 6.5% on LibriSpeech test-other

🔗 Download Russian model: Link to model

For other languages or custom domains, contact me:

📧 otosaku.dsp@gmail.com

⚠️ Input audio must be sampled at 16kHz and normalized to [-1.0, 1.0], strictly 2559 samples per chunk (80ms at 16kHz)

This package is designed with privacy in mind:

• Runs entirely on-device
• No cloud calls or external dependencies

If you have any questions, suggestions, or are interested in adapting the model to another language or domain:

Email: otosaku.dsp@gmail.com

MIT License