Repository structure

• Directoryvt/ the server: one module per responsibility (details in Architecture)

  • app.py entry point: arguments, model loading and HTTP routes
  • httpws.py own HTTP and WebSocket server on asyncio
  • session.py one session per tab: WebSocket protocol and permissions
  • processor.py processing queue and downloads
  • vad.py energy-based voice detector
  • asr.py Whisper on CTranslate2
  • streaming.py streaming recognition with local agreement
  • translate.py translators on CTranslate2
  • textproc.py text before and after translating
  • glossary.py each library’s glossary
  • speakers.py diarization and voice references
  • dub.py per-sentence dubbing
  • audiofx.py own audio code in numpy (WSOLA, resampling, loudness)
  • tts.py voice process client
  • tts_worker.py the voice process
  • ipc.py frames over pipes between processes
  • enhance_df.py reference cleanup with DeepFilterNet
  • resources.py idle mode and GPU turns (and recent timings of each computation)
  • hardware.py what machine this is: GPU, VRAM, RAM, CPU (and a lab’s machine with TRANSLAY_EMULATE)
  • profiles.py performance profiles with the speeds measured in the labs
  • perf.py Preferences › Rendimiento: profile, limits, warnings and restart
  • limits.py RAM and CPU limits with a Windows Job Object
  • trash.py 30-day trash (deleted videos and libraries)
  • jsonio.py JSON that never ends up half-written (temp file + replace, .bak copy)
  • gpu.py CUDA DLLs on Windows and the CTranslate2 cache
  • store.py each video’s transcript
  • library.py videos and collections
  • vaults.py several libraries, each in its own folder
  • prefs.py preferences schema
  • youtube.py downloads with yt-dlp
  • cookies.py YouTube session
  • exporter.py exports with ffmpeg
  • auth.py access and API keys
  • api.py HTTP API v1 and file guard
  • events.py event catalog
  • webhooks.py sending events to other apps
  • repair_refs.py repair of old voice references
  • __init__.py version (__version__)
• Directorydesktop/ the desktop app for Windows (Tauri 2; see Desktop app)

  • Directorysrc-tauri/ the Rust part: window, environment installation (runtime.rs), hidden server (server.rs), tray

    • …
  • Directoryui/ start screens: preparing, opening and error (uses the font, colors and buttons from web/)

    • …
  • Directorylocks/ pinned packages with their hashes: server.txt (server) and voice.txt (voice engine)

    • …
  • tools.json tools bundled in the installer (uv, Python, ffmpeg, Node.js) with their SHA-256
  • scripts/prepare.mjs copies what is shared from web/ and downloads and verifies the tools before building
• Directoryweb/ the interface, no framework and no build step (details in Interface)

  • index.html the single page
  • api.html API reference served by the app
  • Directorycss/ tokens, base, components, shell, views and patterns

    • …
  • Directoryjs/ components, connection, views, menus and routes (desktop.js: the title bar in the desktop app)

    • …
  • Directoryfonts/ Plus Jakarta Sans (OFL), served by Translay itself

    • …
• Directorytools/ measurement and testing tools, no browser (see Testing and evaluation)

  • Directorylab/ hardware labs: VRAM ballast, benchmark and per-machine runs

    • …
  • dub_file.py dubs a whole video and exports <name>.es.mp4
  • eval_tts.py compares voice engines
  • eval_pipeline.py measures recognition and diarization
  • eval_translation.py compares translators
  • transcribe_file.py simulates live playback
  • make_dialogue.py generates test dialogues with Kokoro and noise
  • make_sample.ps1 generates a test video with the Windows voices
  • Directorysamples/ scripts of the test videos

    • …
• Directorytests/

  • test_core.py tests without a GPU
• Directorydocs/ this documentation (Astro + Starlight; see Documentation)

  • astro.config.mjs site and sidebar configuration
  • Directoryscripts/

    • gen_reference.py reads vt/ and generates the reference tables
    • sync.mjs runs gen_reference.py and builds the changelog pages before building
    • check-links.mjs after building, checks every relative link and anchor in both languages
  • Directorysrc/

    • Directorycontent/docs/ the pages in English; es/ the same in Spanish

      • …
    • components/Diagram.astro draw.io-style diagrams (data in src/diagrams.ts, shared by both languages)
    • Directorydata/ JSON generated by gen_reference.py

      • …
• Directorydata/ everything the app stores; not in git (see Data on disk)

  • …
• Directorysamples/ test videos and their ground truth; not in git

  • …
• Directory.venvs/ one environment per voice engine (omnivoice, chatterbox, voxcpm, kokoro); not in git

  • …
• run.py starts the server (python run.py)
• iniciar.bat the same with a double click on Windows; passes arguments to run.py
• requirements.txt server dependencies
• Dockerfile image with CUDA, the server and the voice engine (see Docker)
• docker-compose.yml service with GPU, data/ and the models volume
• .env.example .env template with the owner password for Docker
• .dockerignore what does not go into the image
• .gitignore what does not go into the repository
• README.md project overview in English (README.es.md in Spanish)
• CHANGELOG.md changes of each version in English (CHANGELOG.es.md in Spanish)

Note

The voice engine is not installed with requirements.txt: it lives in .venvs/<engine>/ with its own Python and torch, and the server launches it as a separate process. See From source; the desktop app creates them on its own.