zenKev (Zen Voice Navigator)

Accessible, Local, Deterministic Voice Navigation Engine for Zen Browser (Gecko)

zenKev provides native, zero-latency, CPU-first voice control for Zen Browser through deep integration with Gecko's Accessibility Object Model (AOM), bypassing the visual DOM and communicating with an ultra-fast Rust inference daemon.

---

Architecture Overview

       [ User Voice Command ]
                  │
                  ▼
   [ Audio Pipeline: Silero VAD + Whisper STT ]
                  │ (Transcript text)
                  ▼
   [ Gecko Content: ZenVoiceNavChild ]  ◄── Extracts actionable AOM nodes via nsIAccessibilityService
                  │
                  ▼ (IPC sendQuery)
   [ Gecko Chrome: ZenVoiceNavParent ]
                  │
                  ▼ (Stdio NDJSON)
   [ Rust Daemon: zen-voice-engine ]    ◄── Lexical pruning & Intent Decision (<0.1 ms latency)
                  │
                  ▼ (Target Action)
   [ Gecko Actuator ]                   ──► Direct accNode.doAction(0) / takeFocus()

---

Core Components

1. modules/voice-nav/ (Gecko JSWindowActors):

   • ZenVoiceNavChild.sys.mjs: Runs in content processes. Iteratively traverses nsIAccessible trees, prunes non-interactive elements and elements outside the current viewport.
   • ZenVoiceNavParent.sys.mjs: Runs in the parent/chrome process. Acts as the orchestrator and IPC bridge.
   • ZenVoiceEngineClient.sys.mjs: Manages the lifecycle and Stdio NDJSON pipe to the Rust daemon.

2. modules/voice-engine/ (Rust Inference Daemon):

   • lexical_ranker.rs: Sub-millisecond pre-ranking and pruning using token n-grams and normalized similarity.
   • nli_engine.rs: ModernBERT Natural Language Inference intent decision classifier.
   • mimalloc allocator: Zero-fragmentation memory allocation under high throughput.

---

Screen Reader Coexistence (JAWS / NVDA / VoiceOver)

zenKev does not intercept OS-level accessibility APIs (IAccessible2, UIAutomation, NSAccessibility). It queries Gecko's internal nsIAccessibilityService concurrently as a passive observer, preserving virtual cursors and focus states for external screen readers.

---

Building and Testing

1. Rust Daemon Tests

cd modules/voice-engine
cargo test --release

2. Live Stdio Benchmark

cargo build --release --manifest-path modules/voice-engine/Cargo.toml
node modules/voice-engine/tests/benchmark_throughput.mjs

3. Serving the zenKev Bridge (Kev-4B NLI Model)

The fine-tuned Kev-4B checkpoint is hosted on Hugging Face: Ttotopo27/zenkev-v6.

To start the local HTTP bridge server:

python scripts/serve_zenkev_bridge.py

Note: If no local checkpoint path is defined via ZENKEV_MODEL_PATH, the bridge will automatically download and resolve the checkpoint from Ttotopo27/zenkev-v6.

---

License

Mozilla Public License 2.0 (MPL-2.0). See LICENSE for details.