v1.5 Release

Privox Release Notes

v1.5 (Refiner Quality & Latency Optimization)

Release date: 2026-07-03

This release introduces significant optimizations to the AI refiner engine, dramatically improving prompt-caching efficiency, transcription latency, multilingual formatting, and formatting correctness. We have also added a regression evaluation harness to systematically track quality across runs.

⚡ Refiner Performance & Latency Optimizations

• Static-Prefix KV Cache Reuse: Re-ordered prompt blocks by stability so llama.cpp's automatic longest-common-prefix KV cache is reused on consecutive dictations. This reduces consecutive dictation refinement latency significantly (p95 latency dropped from ~1.9s to under 0.45s).
• Prefill & Stop-Token Restructure: Restructured the chat template to prefill the assistant turn with <refined> and stop generation immediately at </refined>. This guarantees the LLM's output matches the expected tag structure and eliminates prompt echo by design, allowing us to retire ~150 lines of complex extraction heuristics.

🛠️ Deterministic Offloading & Language Support

• Deterministic Hesitation Filler Strip: Extended hesitation filler stripping to Japanese (あのー, えーと etc.), French (euh/heu), German (ähm/äh/mhm), Spanish (leading eh), Korean (leading 어/음), and Chinese (leading 呃/嗯). Ambiguous fillers are stripped utterance-initially, while unambiguous fillers are stripped anywhere.
• CJK Inverse Text Normalization (ITN): Added deterministic conversion of spoken numbers and basic math terms into digits and symbols for Chinese, Japanese, and Korean. (e.g., Chinese spoken math 三加五等於八 becomes 3 + 5 = 8, Japanese relative time 三時 becomes 3時, Korean native numerals before hour counter 세 시 become 3시).
• Language-Switch & Repetition Safeguards: Added a Japanese language-switch guard to prevent the LLM from rewriting Japanese input into Chinese characters when kana is present. Enhanced character-level repetition guards and added automatic retry on empty or degenerate outputs.

🐛 Bug Fixes & Stability

• Fixed Startup Launch Failure: Restored the module-level import logic in src/voice_input.py by relocating the psutil Windows background anti-throttling block, fixing a syntax/compilation error that broke launch from source.
• Fixed Gemma 4 E4B IT Degeneracy: Resolved a critical issue where n_batch=256 caused the llama.cpp engine to generate endless <unused24> tokens with Gemma 4 GGUFs, which previously triggered a silent fallback to raw transcripts. Capped n_batch to 512 (or 128 on lower-VRAM cards).
• Fixed Japanese Translation Corruption: Prevented zhconv from converting Japanese Kanji ideographs during Chinese output script normalization.
• Regression Evaluation Harness: Added scripts/eval_refiner.py and scripts/eval_cases.json to allow developers to run pixi run eval-refiner to verify refiner performance against 25 real-world multi-lingual cases.
• Retired Whisper Turbo Cantonese (CT2) Preset: Wake-up performance improvements make the dedicated Cantonese faster-whisper model unnecessary. It has been removed from Settings; legacy selections migrate to Qwen-ASR v3 1.7B (or 0.6B where folder-id migration applies). Multilingual / Cantonese dictation now uses Qwen-ASR directly.

---

v1.2 Release

Privox Release Notes

v1.2 (GPU Stability & VRAM Fix Update)

Release date: 2026-05-26

This update resolves a crash that affected users with NVIDIA GPUs in the 10–12 GB range (e.g. RTX 3080, 4070, 4070 Ti) when using Qwen-ASR alongside a refiner model. It also fixes a separate issue where VRAM was not fully released during the VRAM Saver idle state.

🐛 Bug Fixes

• Fixed CUDA Out-of-Memory crash on startup (Qwen-ASR + 10–12 GB GPU): When both the grammar/refiner model and the Qwen-ASR model needed to share a mid-range GPU, the refiner could fill all available VRAM before the ASR model had a chance to load. Privox now automatically calculates how much VRAM the ASR model will need and reserves that headroom when loading the refiner — both models coexist without crashing.

• Fixed VRAM not released during idle (VRAM Saver): After a recent update introduced smarter GPU layer placement via device_map, a subtle issue caused VRAM to remain occupied even after the VRAM Saver triggered and the app showed "Idle (VRAM Free)". The root cause was that PyTorch's accelerate dispatch hooks intercept .cpu() calls on device-mapped models, silently preventing them from moving to CPU. Privox now correctly removes these hooks before offloading, ensuring GPU memory is fully freed at idle.

• Fixed VRAM Saver timeout setting not accepting 0 (disabled): The Settings UI now correctly allows setting the VRAM Saver timer to 0 to disable it entirely, instead of clamping the value to a minimum of 5 seconds.

🔧 Improvements

• Smarter VRAM allocation for Qwen-ASR: The ASR model now loads with an explicit VRAM budget (device_map=auto + max_memory) on GPU systems, distributing model layers within available headroom rather than attempting a single large transfer that could fail.
• VRAM flush between model loads: A garbage-collection and CUDA cache flush step was added between the grammar model and ASR model load sequences, ensuring any scratch memory held by the refiner is returned to the driver before ASR initialises.
• Improved ASR cleanup on backend switch: Switching ASR models in Settings now correctly frees GPU memory from the previous model, including any accelerate dispatch hooks.

v1.1 Release

Privox Release Notes

v1.1 (Windows Experience Update)

Release date: 2026-05-15

This update focuses on making Privox faster, more reliable, and easier to use on Windows. We've combined several recent improvements into this single "v1.1" release to ensure the best experience for all users.

🚀 Performance & Speed

• Lightning Fast AI: We've optimized the app to take full advantage of modern NVIDIA graphics cards (including the RTX 40 and 50 series). This allows the AI to process your speech almost instantly without slowing down your computer.
• Smoother Transcription: The system is now much more efficient at turning your voice into text, providing a snappier feel especially on newer hardware.

🛠️ Reliability & "Just Works"

• Fixed Startup Issues: We resolved a common technical problem that caused the app to occasionally fail to start on certain Windows systems. It should now launch reliably every time.
• Better Typing & Pasting: When you finish speaking, Privox is now smarter about finding your active window and pasting your text exactly where your cursor is.
• Modern AI Engine: Upgraded core transcription engine (llama-cpp-python) to v0.3.23, enabling official Windows binary support and faster, more stable AI operations.
• Cleaner Installation: The app now does a better job of keeping your computer tidy by automatically cleaning up temporary background files.
• Refined Installer UI: Unified fonts and styling across the installer, and implemented correct taskbar icon association for a more professional look.
• Improved Hotkey Reliability: Implemented physical key-state verification to prevent the hotkey from getting "stuck" or missed during high system load.
• Pre-warm Models: Added a new "Pre-warm Models on Startup" feature (enabled by default) so the AI is ready to transcribe instantly without a first-use delay.

🛡️ Privacy & Security

• Strictly Local: We've further enhanced our privacy protections to ensure that your dictation never leaves your computer.
• Silent & Unobtrusive: The app now runs more smoothly in the background, keeping itself out of your way while you work.

Privox v1.0.0 - Initial Release

Privox v1.0.0 - Initial Release

Privox is a private, local-first voice input assistant for Windows that transcribes your speech and intelligently refines it using state-of-the-art AI models.

✨ Key Features

🎙️ High-Precision Transcription

• Faster-Whisper Engine: Blazing fast transcription using the latest AI models.
• True Multilingual Support: Specialized logic for English, Cantonese, Traditional Chinese, Japanese, Korean, and more.
• Intelligent LID (Language Detection): Automatically identifies the language you are speaking with high-confidence safety thresholds.
• Optimized Performance: Parallel model loading and optimized wake-up sequences ensure minimal latency when starting or resuming from idle.

✍️ Intelligent Refinement

• Smart Refiner: Uses Llama-3.2-3B to polish your speech into grammatically perfect text while removing filler words like "uh" and "um".
• Hallucination Safeguards: Multi-layered protection prevents the AI from adding unintended text or "hallucinating" on short inputs.
• Anti-Assistant Protection: Persona guards ensure the AI stays focused on refining your text rather than acting as a chatbot.
• Clean Output: Automatic meta-commentary stripping removes trailing AI "notes" or internal explanations.
• Custom Personas: Choose how Privox writes—act like a Technical Writer, Engineer, or Lawyer with a single click.

�️ Private & Secure

• 100% Local: All processing happens on your computer. Your privacy is guaranteed.
• Privacy-First Logging: User text and refined output are never written to logs—only diagnostic metadata is tracked.
• Legitimacy Metadata: Embedded Windows Version Info and application manifests reduce false-positive flags from antivirus software.
• Smart Auto-Stop: Advanced silence detection and safety timers ensure the microphone stops recording when you do.

✍️ Premium Experience

• Liquid Glass UI: A clean, monotone aesthetic with centered progress updates and professional typography.
• One-Key Magic: Single hotkey operation for seamless start/stop recording.

🛠️ Effortless Local Setup

• Smart Installer (bootstrap.py): Automatically manages portable Python environments and GPU/CUDA dependencies.
• GPU Orchestration: Built-in support for NVIDIA GPUs with automatic fallback and repair for CPU mode.
• Disk Safety: Verifies space and system requirements before downloading large AI assets.

🖥️ System Requirements

• OS: Windows 10/11.
• GPU: NVIDIA GPU (CUDA 12+) recommended; CPU support available,.
• Disk: ~15GB free space for the high-quality local AI "brains".