WebUI - Initial native Svelte WebUI embedded inside audiocpp_server.exe — no Python, Node.js, or Gradio needed

[mirek190]

Summary

This introduces the initial native SvelteKit/TypeScript WebUI and embeds its complete static output inside audiocpp_server.exe.

Normal browser inference now needs only the server executable plus the selected CPU/GPU runtime libraries. Python, Node.js, and Gradio are not runtime dependencies:

Browser
   |
   | HTTP on localhost
   v
audiocpp_server.exe
   |-- compiled-in Svelte HTML/CSS/JavaScript
   |-- native /v1 server APIs
   `-- audio.cpp model sessions
          |-- CPU
          `-- CUDA / other configured backends

Node.js is used only when a developer rebuilds the frontend artifact. Python is used only when the user explicitly asks the optional model installer to download or convert a package; loading a prepared model and running inference remain native. The existing Gradio UI is retained as a compatibility option but is not used by this UI.

What is included

• Build webui/native as a single static document and convert it to a generated C++ byte array during CMake configuration.
• Serve the compiled asset at / and /index.html, with cache, CSP, and content-type safety headers.
• Add server switches --ui, --no-ui, and --ui-management; --ui can start a model-management host without an external server JSON file.
• Add guarded native endpoints for dynamic model loading/unloading, path checks, temporary uploads, and asynchronous package-install status.
• Cover the primary audio.cpp workflows:
  • text to speech and voice cloning
  • ASR/transcription and near-live ASR
  • music generation
  • voice conversion
  • source separation, including HTDemucs, BS-RoFormer, and Mel-Band RoFormer presets
  • audio analysis
  • voice design
• Add long-text splitting/merge, microphone capture, multiple-output playback, timing, model catalog/settings, and a browser-local IndexedDB voice library.
• Route package-spec installs through model_manager_v2.py; keep converter-input packages on the deprecated helper until those packages migrate.
• Extend portable Windows packaging with the optional model-management resources.
• Document native and legacy WebUI setup and add focused configuration/model-installer tests.

Build and test commands

Frontend:

cd webui/native
npm ci
npm run check
npm run build

CPU server and tests:

cmake --build build/native-ui-cpu --config Release --target audiocpp_server server_config_test server_model_installer_test --parallel 8
ctest --test-dir build/native-ui-cpu -C Release --output-on-failure -R "server_(config|model_installer)_test"

CUDA server (RTX 3090 / compute capability 8.6, CUDA 12.4):

.\scripts\build_windows.ps1 -Preset windows-cuda-release -Target audiocpp_server -CudaArchitectures 86 -Jobs 8

Run the native UI:

.\build\windows-cuda-release\bin\audiocpp_server.exe --ui --backend cuda --host 127.0.0.1 --port 18991

Then open http://127.0.0.1:18991/.

Results

• svelte-check: 0 errors, 0 warnings.
• Frontend static build: passed.
• CPU audiocpp_server: built successfully against current main (238ab6a).
• server_config_test: passed.
• server_model_installer_test: passed.
• CUDA audiocpp_server: built successfully with CUDA 12.4 for SM 8.6.
• CUDA smoke test:
  • GET /health: status=ok, backend=cuda, ui=true, ui_management=true
  • GET /: HTTP 200, 167,097 bytes
  • embedded page contained the current BS-RoFormer catalog entry

Generated artifacts:

• webui/native/dist/index.html
• build/native-ui-cpu/bin/Release/audiocpp_server.exe
• build/windows-cuda-release/bin/audiocpp_server.exe

Known limitations

• Model download/conversion from the WebUI is optional and still invokes the repository Python model-manager helper. Users with already prepared models do not need Python.
• The legacy Gradio interface remains in the repository for advanced Python-assisted or compatibility workflows; it is separate from the embedded native UI.
• Near-live ASR uses repeated server requests rather than a browser-side model runtime.

[0xShug0]

The UI looks super cool!

[mirek190]

Pushed commit 684832b with the current native WebUI model-library update.

What changed

• Added asynchronous model-package installation with live byte and percentage progress in the embedded UI.
• Exposed GGUF Q8, GGUF BF16/FP16, and Safetensors choices with their package-specific download sizes and availability state.
• Detects already-installed package variants using their declared files, disables the corresponding button, and shows Already downloaded.
• Allows precision variants that share a package directory to coexist without treating the directory itself as a conflict.
• Added per-variant trash controls and a server uninstall endpoint. Removal deletes only files declared by that package and preserves sibling variants and unrelated files.
• Corrected Pocket TTS package paths and its gated Safetensors messaging.
• Made model cards, Open controls, and package-choice rows consistent in size; same-row cards now align with uniform vertical gaps.
• Updated the embedded production WebUI artifact and documentation.

Validation

• npm --prefix webui/native run check ? 0 errors, 0 warnings
• python -m unittest webui.test_model_manager_webui ? 16 tests passed
• server_model_installer_test ? passed
• CPU audiocpp_server Release build ? passed
• CUDA audiocpp_server Release build ? passed
• Browser layout check ? matching card heights within each row and consistent 12 px row gaps

[mirek190]

Pushed follow-up commit 55f495a

Root cause

All ModelInstaller instances used the same %TEMP%/audiocpp-model-installer directory and reused names such as package-sizes-1.json. A test process or second server could overwrite the live server's inventory snapshot. The UI would then lose the real installed-package records and leave an existing package clickable.

Fix

• Give every installer/server instance a unique temporary job directory.
• Remove that private directory when its installer state is released after normal shutdown.
• Treat a package whose complete declared file set already exists as an idempotent successful install.
• Continue rejecting genuinely partial packages unless --overwrite is requested.
• Add regression coverage proving two installer instances cannot overwrite each other's inventory output.

Validation

• Python model-manager suite: 17 tests passed.
• C++ server_model_installer_test: passed, including cross-instance isolation.
• CPU and CUDA Release server builds: passed.
• Live CUDA inventory: 127 packages returned; Pocket English Q8 and BF16 both reported installed: true.
• Exact Pocket Q8 install API request against the existing file: completed with exit code 0 and already installed, with no download or overwrite.

[mirek190]

Pushed follow-up commit 0fe5bca fixing installed-package precision selection and stale inventory after downloads.

Changes

• Installed Q8/BF16/FP16/Safetensors buttons remain enabled and now select the model variant instead of trying to download it again.
• Buttons retain the format label and distinguish Selected ? size from Downloaded ? size.
• Open uses the selected installed package path.
• A successful install invalidates the server's cached package inventory.
• The frontend refreshes that inventory once for each newly completed install, so Studio ? Models navigation cannot restore stale installed flags.
• Added server regression coverage for inventory invalidation after a successful preparation.

Validation

• Svelte check: 0 errors and 0 warnings.
• C++ server_model_installer_test: passed.
• CPU and CUDA Release server builds: passed.
• Browser test with Pocket TTS:
  • Q8 and BF16 both remained marked downloaded.
  • Switched Q8 ? BF16, navigated Models ? Studio ? Models, and BF16 remained selected.
  • Switched back to Q8 and pressed Open.
  • Studio resolved models/PocketTTS-GGUF/english/pocket-tts-english-q8_0.gguf.