Changelog - RocketWhisper | AI Speech Recognition & Transcription

[Fix] Speech recognition stops working after resuming from hibernate or sleep — On some GPUs (such as AMD), recognition could stay unavailable until the app was restarted. RocketWhisper now restarts itself automatically on resume and after a GPU error so the GPU is initialized fresh. Text in the main window is carried over.
[Fix] Groq AI processing works again — Groq retired LLaMA 3.3 70B / 3.1 8B on August 16, 2026; RocketWhisper now uses GPT-OSS 120B / 20B.
[Fix] Claude Sonnet 5 / Opus 5 could return an empty result
[Fix] After a failed recognition, the previous result could be pasted again or used for an AI command
[AI Processing] Updated models — OpenAI GPT-6.1 Sol / GPT-6 Luna, Claude Sonnet 5.5 / Opus 5.5, Gemini 3.8 Flash (default). Previously selected models move to a successor in the same price range automatically.
[Support] Diagnostic logs — Makes it easier to investigate problems you report (saved in %APPDATA%\RocketWhisper\logs; audio, recognized text and API keys are never recorded).

[Accuracy] Hallucination suppression with VAD — Built-in Silero voice activity detection skips silent or noise-only segments before recognition, preventing fabricated sentences from silent recordings (toggle on the Model & Language tab, on by default).
[GPU] CUDA 12 runtime added — CUDA acceleration now also works on NVIDIA systems that have the CUDA Toolkit 12 installed (CUDA acceleration requires the NVIDIA CUDA Toolkit 13 or 12; see Help > Troubleshooting). Vulkan (NVIDIA/AMD/Intel) continues to work as before. The About dialog shows which recognition engine is in use.
[AI Processing] Refreshed model list — OpenAI GPT-5.6 (Terra/Luna/Sol), Anthropic Claude Sonnet 5 / Haiku 4.5 / Opus 5, Google Gemini 3.7 Flash and more. Previously selected legacy models fall back to the provider default automatically.
[Distribution] Code-signed installers — Installers are now digitally signed by "Mojosoft Co., Ltd.".
[Internal] Speech engine updated to Whisper.net 1.9.1 — Unused components removed for a smaller install.

[Stability] Auto-recovery from GPU crashes — Automatically reloads the Whisper model when CUDA/Vulkan GPU driver crashes after hibernate/sleep resume.
[Stability] Transcription timeout — No longer hangs indefinitely on GPU freeze. Detects timeout and auto-reloads the model.
[Stability] Power resume retry — Retries model reinitialization up to 3 times when driver recovery is slow.

Transcription speed display — Shows processing time and RTF (Real-Time Factor) in the status bar after recognition completes. Instantly see your transcription performance.

[Stability] Auto-recovery after hibernate/sleep resume — Automatically reinitializes Whisper model and audio devices after system resume. Detects and recovers from CUDA GPU memory invalidation and audio driver inconsistencies.
[Robustness] Graceful error handling — Improved handling of unexpected errors to prevent the app from becoming unresponsive. Displays error message and exits gracefully.
[Robustness] Audio device enumeration protection — App no longer crashes when audio driver errors occur during microphone enumeration.

USB microphone hotplug detection — Real-time detection of microphone connect/disconnect with automatic device list refresh.
Video file transcription — Transcribe directly from MP4/MKV/AVI/MOV/WebM video files using FFmpeg audio extraction.

Custom Instructions — Execute specific AI instructions instantly with dedicated hotkeys. Fast processing for translation, summarization, proofreading without voice input.

[Fix] The Right Alt key could get stuck in the pressed state (recurrence)
[Improvement] New notification sounds — Recording start/stop sounds now use subtle standard Windows sounds.

[New] Recording start/stop sounds — A sound tells you when recording starts and stops (can be turned on or off in Settings).
[Improvement] Smaller recording indicator
[Fix] The Right Alt key could get stuck in the pressed state — Reworked key handling for Right Alt hold-to-talk mode.
[Improvement] Right Alt works normally when recording is unavailable — While the model is loading or recognition is in progress, the Right Alt key is passed through to other apps as usual.

[New] Right Alt hold-to-talk mode — Records only while you hold the Right Alt key. Double-tap it to keep recording until you tap it again.
[New] AI Command Mode — Select text, press the dedicated hotkey and give a voice instruction; the AI processes the selected text as instructed.
[New] Voice Search — Ask by voice to look something up, and RocketWhisper runs a Google search in your browser.
[Change] New default hotkeys — Recording: Right Alt (hold); AI Command: Ctrl + Space.
[Change] "AI Formatting" renamed to "AI Processing"

[GPU] NVIDIA CUDA acceleration — CUDA acceleration requires the NVIDIA CUDA Toolkit to be installed (for the current requirements, see Help > Troubleshooting).
[Improvement] Default model changed to large-v3-turbo — Faster recognition while keeping high accuracy.
[Improvement] Resumable model downloads — Detects incomplete model downloads and resumes them where they left off.
[AI Processing] Anthropic Claude models updated to the 4.5 generation
[Fix] Clipboard issue with auto-paste
[Improvement] Improved misrecognition correction rules
[Fix] Error in the correction rules window