Whisperlocal

Parley

An iOS app that records conversations, identifies speakers, transcribes them, and summarizes them — entirely on your device. No cloud, no analytics, no telemetry.

Originally prototyped under the name Whisperlocal. Renamed to Parley — an English word for a private confidential conversation.

Status: Phase 4. End-to-end on-device pipeline: record → speaker diarization (SpeakerKit) → transcription (WhisperKit) → summarization (Apple Foundation Models). Braun-inspired UI. Persistent recording library. Share-sheet export.

Privacy posture

• NSAppTransportSecurity blocks arbitrary loads (HTTPS-only).
• No iCloud / CloudKit containers; no analytics SDKs.
• Microphone usage string is explicit about local-only processing.
• One network call, ever: the first launch lets WhisperKit fetch the Whisper CoreML model from Hugging Face. After that, audio and transcripts never leave the device.

Requirements

• macOS with Xcode 15.3+
• iOS 17.0+ device or simulator (real device strongly recommended — Whisper’s CoreML pipeline is unreliable on the iOS Simulator)
• XcodeGen — brew install xcodegen

Getting started

brew install xcodegen        # one-time
xcodegen generate            # produces Parley.xcodeproj
open Parley.xcodeproj

In Xcode: pick your iPhone in the destination picker, then ▶ Run. First launch downloads the on-device speech and speaker models in the background; you’ll see “Preparing models…” until they’re ready.

Project layout

Parley/
├── ParleyApp.swift                # @main entry
├── SessionStore.swift             # state machine: record → transcribe → summarize
├── Info.plist
├── Parley.entitlements
├── Audio/
│   └── AudioRecorder.swift
├── Transcription/
│   ├── AudioFileReader.swift
│   └── WhisperKitTranscriptionService.swift
├── Summarization/
│   └── SummarizationService.swift
├── Models/
│   ├── Recording.swift
│   ├── RecordingStore.swift       # persistent library
│   └── RecordingExport.swift      # share-sheet formatting
└── Views/
    ├── DesignSystem.swift         # Braun 1968 palette + type
    ├── RootView.swift
    ├── LibraryView.swift
    └── RecordingDetailView.swift

Roadmap

• Phase 1 (done): Scaffold + recording + mock services + UI flow.
• Phase 2 (done): WhisperKit for on-device transcription on the Neural Engine.
• Phase 3 (done): SpeakerKit diarization, Apple Foundation Models summarization, share sheet, date/time on results.
• Phase 4 (done): Braun 1968 redesign, persistent recording library, formatted share output.
• Phase 5 ideas: Recording playback, on-device search, accessibility audit, app icon + launch screen, multilingual model option.

Apple Intelligence requirement for summarization

The on-device summarizer uses Apple’s FoundationModels framework, which needs iOS 26+ running on an Apple Intelligence capable device (iPhone 15 Pro / Pro Max / 16 family or newer). On any other device, summarization gracefully degrades to a placeholder — transcription and diarization still work everywhere iOS 17+ runs.

Choosing a different Whisper model

WhisperKit defaults to openai_whisper-base.en (small, fast, English-only). To use a different one — e.g. the more accurate openai_whisper-small.en or multilingual openai_whisper-base — change defaultWhisperModel in Parley/Transcription/WhisperKitTranscriptionService.swift. Available models are listed in WhisperKit’s README.

License

TBD — pick one before publishing.