xTranscript v1.1.0

A free, open-source (MIT) native macOS transcription tool that sends audio to xAI's Speech-to-Text API and writes .txt / .srt files with word-level timestamps.

Features

• Drag-and-drop GUI — multi-file queue with per-file status, retry, overwrite protection
• 11 languages via grok-stt (English, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean — xAI's streaming endpoint requires an explicit language)
• Universal input format support — every file is decoded by the embedded LGPL-only FFmpeg build (MP3, M4A, MP4, MOV, WAV, FLAC, AAC, AIFF, CAF, ALAC, MKV, WebM, OGG/Opus, AVI, WMV, WMA, …)
• Dual output — .txt transcript and .srt subtitles
• Real-time progress — progress bar and live log during transcription
• Persistent settings — API key, output folder, formats, language saved across launches
• No limits, no accounts, no in-app purchases — bring your own xAI API key; xAI bills the audio minutes directly to your account

Requirements

• macOS 14+
• Xcode 16+ (for building)
• An xAI API key (get one at console.x.ai)
• Internet connection (every file is uploaded to xAI)

Build & Run

Open Transcript.xcodeproj in Xcode, select the xTranscript scheme, and hit Run. Paste your xAI API key into the Settings sidebar.

Architecture

Transcript/                         # SwiftUI app
├── TranscriptApp.swift             # App entry point
├── Models.swift                    # Shared types (TranscriptionSettings, LabeledSegment, TranscriptLanguage, errors)
├── TranscriptionViewModel.swift    # UI state management, file queue, retry, sandbox folder grants
├── ContentView.swift               # Three-column layout (sidebar, log, file queue)
├── SidebarView.swift               # Settings panel (output, language, formats, API key, acknowledgements)
├── TranscriptionService.swift      # Orchestrator: open PCM stream → pipe through WebSocket → group → write
├── AudioExtractor.swift            # Thin entry point: opens an FFmpegPCMReader, probes duration
├── FFmpegPCMReader.swift           # LGPL FFmpeg decoder for every supported format
├── XAIClient.swift                 # Streaming client for wss://api.x.ai/v1/stt
├── OutputGenerator.swift           # TXT and SRT generation
└── Resources/LICENSES.txt          # Bundled acknowledgements (shown in-app)

Vendor/
└── FFmpeg.xcframework              # ~7.7 MB universal static lib — built by scripts/build-ffmpeg.sh

TranscriptTests/                    # XCTest: output formatting + FFmpeg decoder fixtures

How it works

1. Files are dropped onto the right panel (multiple files supported, processed sequentially)
2. Existing output files are detected — user is prompted before overwriting
3. Audio is decoded to 16 kHz mono Int16 PCM via the embedded LGPL FFmpeg build (Vendor/FFmpeg.xcframework) and streamed in ~250 ms chunks
4. Each chunk is pushed over a WebSocket to wss://api.x.ai/v1/stt with the chosen language code; chunk-final partials stream back in real time
5. On transcript.done the chunk-final partials are mapped to LabeledSegments (one per ~3 s server chunk)
6. .txt and .srt files are written next to the input file (or to a custom folder)

Speaker diarization is wired end to end but currently forced off: xAI's streaming endpoint runs out of memory with diarize=true on inputs longer than about a minute (reported upstream, 2026-04-22). The toggle in SidebarView is commented out until xAI ships a fix.

Tests

18 unit tests cover the text/SRT formatting (TranscriptTests/OutputGeneratorTests.swift) and the FFmpeg decoder on small MKV/WebM/OGG/MP3 fixtures (TranscriptTests/FFmpegPCMReaderTests.swift). Run them with Cmd+U in Xcode, or:

xcodebuild -project Transcript.xcodeproj -scheme xTranscript -destination 'platform=macOS' test

Rebuilding FFmpeg

The Vendor/FFmpeg.xcframework is checked into the repo (~7.7 MB). To rebuild it from source — for an FFmpeg version bump or a CVE refresh — run:

./scripts/build-ffmpeg.sh --clean

The script clones the pinned FFmpeg tag (see scripts/build-ffmpeg.sh for the version), configures it with --disable-everything and an explicit allow-list of audio demuxers/decoders/parsers (no GPL or non-free codecs, no network/protocol code), builds for arm64 + x86_64, and packages the result as a static XCFramework. Verifies post-build that no x264/x265/fdk/amr/gsm symbols leaked in.

Distribution

The app is configured for Mac App Store submission:

• App Sandbox is enabled (Transcript/xTranscript.entitlements) with network.client (for the xAI API) and files.user-selected.read-write (for dropped inputs and chosen output folders).
• Privacy Manifest at Transcript/PrivacyInfo.xcprivacy declares audio data collection with purpose = app functionality, and required-reason API use for UserDefaults and file metadata reads.
• Bundle ID is net.eric-nicolas.xtranscript (reverse-DNS of eric-nicolas.net).
• Encryption export compliance: ITSAppUsesNonExemptEncryption = false in Info.plist (system TLS only, no proprietary crypto).
• Privacy policy lives at docs/privacy.md and is served directly by GitHub — no separate web hosting needed. Use these URLs in App Store Connect:
  • Privacy policy URL: https://github.com/enicolas72/transcript/blob/main/docs/privacy.md
  • Support URL: https://github.com/enicolas72/transcript/issues

To archive for submission: Xcode → Product → Archive → Distribute App → App Store Connect. Requires an Apple Developer Program membership and a matching Mac App Distribution certificate.

License

MIT — see LICENSE. The embedded FFmpeg build is LGPL-2.1+; its notice and source-availability pointer are in Transcript/Resources/LICENSES.txt and shown in-app under Acknowledgements.