feat(voicechat): add Windows RTX desktop example

[yifeif-nv]

Background

Add a buildable Windows voice application that demonstrates Nemotron VoiceChat through TensorRT-RTX, with live transcripts, reactive visuals, and a presentation mode. The existing full-duplex example uses Linux/ALSA. This source example also addresses delayed playback and repeated replies observed during longer interactive conversations.

This draft depends on #1218. Its existing commit remains in the comparison to main; review the Windows example commit separately from that prerequisite.

Exit Criteria

• Users can download dependencies and build the Windows desktop application, native bridge, and RTX model locally using documented PowerShell stages.
• The example distributes source, scripts, tests, and documentation only. No third-party binaries, libraries, SDK headers, Python/npm environments, model weights, bundles, or audio/media assets are added.
• Capture, playback, interruption, and context refresh remain bounded during continued conversation and recover without reinjecting abandoned assistant replies.
• Source checks and supported-host validation are recorded separately from uncompleted clean-machine qualification.

Implementation

• Add examples/windows_voicechat: sandboxed Electron UI, canvas/SVG visuals, bounded NDJSON C++ bridge, settings, local diagnostics, stream mode, and labeled model-free rehearsal.
• Add PowerShell setup/run entry points. Setup downloads verified Python, Electron, NVIDIA SDK components, and build tools into an external workspace, installs pinned Python packages, downloads a pinned checkpoint, and builds locally. Relative settings and launchers support relocated local installations. Native setup requires administrator rights only when the Microsoft toolchain is missing.
• Add a source-distribution audit and Windows CI for index contents, JavaScript units, and PowerShell parsing/planning. There is no binary release/archive publication step.
• Make standard TensorRT optional in CMake, select only the Nemotron family for this Windows build, and support Windows DLL discovery and Unicode bundle/cache paths. The example builds and requires trt_rtx.
• Use 20 ms microphone packets to avoid racing the native fallback clock. Refresh carries only the latest unanswered request, with bounded repetition detection and one retry. A public, family-owned live-control interface resets dialogue through a worker barrier while keeping microphone/audio frontend continuity; shared core remains model-agnostic.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

• cmake --build build-windows-rtx --parallel 8, ctest --test-dir build-windows-rtx --output-on-failure, and cmake --install build-windows-rtx: passed after the public-header changes, with 10/10 CTests including real RTX inference. Installed public header and exported interface target verified.
• python -m pytest examples/windows_voicechat/native/test_native_startup.py -q with the staged bridge: 4 passed.
• python -m pytest -q -p no:cacheprovider tools/tests/test_architecture.py tools/tests/test_family_impact.py tools/tests/test_public_source_hygiene.py tools/tests/test_pr_metadata.py: 82 passed.
• node --test examples/windows_voicechat/desktop/tests/*.test.js: 23 passed, including fresh-workspace discovery and relocation.
• python examples/windows_voicechat/audit_source.py --staged: passed for all 40 example source/text files; negative tests in isolated Git repositories rejected force-added binaries, vendored source, oversized files, and encoded media.
• Windows PowerShell 5.1 parser and Setup.ps1 -Plan: passed. An isolated Setup.ps1 -Stage Python downloaded and verified CPython, created a venv, and imported SSL. Spaces/apostrophes in paths, relative launcher relocation, Run.ps1 -ValidateOnly, and App-only rehearsal validation passed.
• node desktop/tests/electron-audio.integration.cjs and node desktop/tests/electron-transcript.integration.cjs from the example directory: passed using downloaded Electron and separately installed Playwright; real WebAudio/IPC, 20 ms capture, interruption/reset controls, and 1,034 events with 300 retained rows were checked.
• python -m pytest -q -p no:cacheprovider families/nemotron_voicechat/tests/test_build_policy.py families/nemotron_voicechat/tests/test_quantization.py: 18 passed.
• Repository-pinned ruff check --config ruff.toml on changed Python source, clang-format --dry-run --Werror on changed C++ source, and git diff --check: passed.
• Full RTX conversation evidence, before the source-packaging/path/public-header refinements: 540.028 seconds, 12 topic checks, six actual context refreshes, spoken and manual interruption, no runtime/renderer errors, and maximum playback lead 0.941333 seconds. Follow-up model-free checks cover the packaging/path refinements. See the example's VALIDATION.md for scope and reproduction.

Hardware, Environment, and Revisions

• Published source head: e803846cfdfbd7dedcbcbb9d23649e5870fbd8ab; its tree is byte-identical to the locally validated source tree cb3410412ec14537403a0297af948e02f67ac56e.
• Windows x64; RTX 5090 with 32 GB VRAM; 128 GB host RAM; NVIDIA driver 591.86.
• MSVC 19.44, CUDA 13.4, TensorRT-RTX 1.6.1.120, CPython 3.12, Electron 44.3.0.
• Prerequisite feat(voicechat): stabilize compressed full-duplex sessions #1218 revision: 7458623963a038fa1ae1f1bac28eee6a5c792514.
• Checkpoint nvidia/NVIDIA-NemotronLabs-VoiceChat-11B revision 359ada7b1c60851e40ff08065f9b0340244f27e0; tokenizer revision 6533e8de2c68e4536bf7c411d7a3ce5734111476.
• TensorRT-RTX bundle uses explicit Thinker W8A8 / higher-precision sensitive projections / FP16 TTS policy. The locally built bundle is approximately 18 GB and is not included.

Not Run / Remaining Gaps

• Full five-stage setup on a freshly installed Windows machine, lower-memory GPUs, other GPU/driver combinations, acoustic echo cancellation, and heavy concurrent GPU workloads remain unqualified.
• A full-model soak was not repeated after source-only packaging, path portability, and public-header relocation. Nine minutes is the measured run, not a claim of empirically proven unlimited operation.
• The broader Linux CI orchestration tests were attempted locally on Windows. Their Bash/compiler/POSIX execution cases cannot establish a passing Linux gate here; remote CI remains authoritative. No test criteria were weakened.
• Remote CI and the automated premerge gate are not claimed as passed by the local checks.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

• Merge feat(voicechat): stabilize compressed full-duplex sessions #1218 first, then rebase this draft to remove its prerequisite commit from the comparison. Do not merge this draft as independent validation of feat(voicechat): stabilize compressed full-duplex sessions #1218.
• Review setup/distribution boundaries and the README first, then the desktop/bridge protocol, shared Windows build/loading mechanics, and family-owned recovery/public live-control interface.
• Third parties are downloaded from upstream on the user's machine. Local assembly includes installed runtimes and their notices; it must not be published as this example's source archive. Upstream repository files already present before this PR are unchanged by that policy.
• Model graph/runtime-version changes require rebuilding the bundle. MSVC's signed bootstrapper and transitive pip resolution are disclosed; this is not a fully locked reproducible binary distribution.

Risk level

• Low
• Medium
• High

The example is opt-in and keeps model logic family-local, but adds Windows native loading/build support and changes experimental VoiceChat conversation recovery. Fresh-machine qualification and prerequisite integration remain necessary before treating it as a turnkey release.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}