Adds local English streaming recognition and adaptive Japanese reading.
- English STT: NVIDIA Nemotron Speech Streaming English 0.6B with pinned NeMo-Speech.cpp 0.3.0 CPU/Vulkan runtimes. Uses continuous 160ms PCM, transient English drafts, and final-only translation/timeline persistence. Qwen and Whisper remain available for multilingual speech.
- Reading: native Irodori speed control, bounded live catch-up, manual-reading base rate, and memory-only generated audio.
- Setup: optional verified/resumable downloads, separate model/runtime storage, explicit English-only behavior, and retained model/runtime licensing notices.
On one RTX 4090 PC, the same 10-second FRB excerpt took 363ms on average with Nemotron Vulkan/560ms context, compared with 1.150 seconds for Qwen CUDA. These are three-run resident-model transcription timings, excluding capture, loading and translation. Streaming first-draft observations were 1.3-1.7 seconds. Accuracy superiority has not been evaluated.
Validation: 90 Rust tests, strict all-target Clippy, nine Python tests, PowerShell 5.1 setup plans, four native CPU/Vulkan/context combinations, Unicode-path fresh ZIP extraction, all entry hashes, static CRT checks and ten headless UI renders.
Preview limitations: new Nemotron GUI operation, region-selection mouse interaction and live adaptive-TTS demo capture remain unverified. See docs/VALIDATION.md for the measured scope. This is not a clean-Windows installation test.
Extract the ZIP to a new folder, launch 地元翻訳くん.cmd, then choose Nemotron in Settings > Speech recognition to prepare its optional data. Existing speech-backend preferences are preserved. The ZIP contains no weights, downloaded runtimes, settings, connection keys, logs or private recordings. Model and external runtime licenses are separate from the application's MIT license.