Building VoiceJot: Low-Latency On-Device Speech Transcription and AI Meeting Minutes in Swift

The Bottleneck: Managing Voice Memos

Voice memos are the most natural way to capture ideas during client meetings, team standups, and college lectures. However, finding actionable information inside raw 30-minute audio recordings is a huge productivity drain.

To solve this, we engineered VoiceJot: Voice to Text, a privacy-first iOS application that combines Apple's native speech recognition with real-time neural formatting and structured AI summaries.

Architectural Highlights

1. AVAudioEngine & SFSpeechRecognizer Pipeline

VoiceJot taps raw PCM audio buffers at 44.1 kHz using an AVAudioNodeTap:

• Partial results are streamed in real time to give users an instant visual typewriter animation.
• Works across 50+ languages with automatic punctuation and capitalization restoration.
• Supports direct audio demuxing from WhatsApp voice messages and video clips via AVAssetReader.

let recognitionRequest = SFSpeechAudioBufferRecognitionRequest()
recognitionRequest.shouldReportPartialResults = true
recognitionRequest.addsPunctuation = true

inputNode.installTap(onBus: 0, bufferSize: 1024, format: recordingFormat) { buffer, _ in
    recognitionRequest.append(buffer)
}

2. Structured Executive Summaries

Instead of dumping raw unformatted text walls, VoiceJot parses transcripts into:

1. Executive Overview: 2-3 sentence core meeting synthesis.
2. Key Discussion Points: Bulleted major arguments.
3. Action Items (To-Do): Direct delegated tasks with deadlines.

Availability & Resources

• App Store Link: https://apps.apple.com/app/id6799866105
• Target OS: iOS 16.0+ Universal

How are you integrating on-device speech transcription in your mobile apps? Let's discuss in the comments below!