Aura v1.0.10

What's Changed

• feat: support Windows/Super and modifier-only hotkeys by @Decidetto in #2
• fix: install required CUDA and cuDNN runtime dependencies by @Decidetto in #3
• fix: localize tray menu and first-launch language by @Decidetto in #4

New Contributors

• @Decidetto made their first contribution in #2

Full Changelog: v1.0.9...v1.0.10

Aura v1.0.9

Full Changelog: v1.0.8...v1.0.9

Aura v1.0.8

What's Changed:

• Resident Whisper Server on CUDA: Introduced resident background server execution (whisper-server.exe) in src-tauri/src/whisper_runner.rs alongside Parakeet's resident server. When NVIDIA CUDA acceleration is active, the server keeps model weights (e.g. large-v3-turbo-q5_0) loaded directly in GPU VRAM (CUDA0), dropping transcription latency from ~15s cold starts down to ~250ms on discrete GPUs.
• Unified On-Demand CUDA Package Downloader: Extended download_gpu_binaries_inner in src-tauri/src/lib.rs to download, SHA-256 verify, and extract both the Sherpa-ONNX CUDA runtime (sherpa-onnx-v1.13.4-win-x64-cuda.tar.bz2) and Whisper cuBLAS runtime (whisper-cublas-11.8.0-bin-x64.zip containing ggml-cuda.dll) into %LOCALAPPDATA%\com.aura.app\binaries\cuda\bin\. This enables one-click GPU acceleration for both engines while keeping the base installer compact (~20–25 MB).
• Dynamic Working Directory DLL Routing: Configured start_whisper_server_unlocked and run_local_whisper in src-tauri/src/whisper_runner.rs to set their process working directory and PATH to %LOCALAPPDATA%\com.aura.app\binaries\cuda\bin\ when CUDA is selected. This allows ggml.dll's ggml_backend_load_all() to dynamically detect ggml-cuda.dll and initialize CUDA0 without requiring DLL copies to system directories.
• Model Verification Spinner & State Feedback: Added a dedicated status field to DownloadProgress in src-tauri/src/whisper_runner.rs and updated src/main.js to transition download cards into an animated installing state with a circular spinner during post-download SHA-256 integrity verification, eliminating the static 5–10s freeze at 100% on large model files.
• Native Neural Punctuation Transition: Removed legacy CT-Transformer external dependency and heuristic string manipulation (apply_voice_punctuation) in src-tauri/src/whisper_runner.rs in favor of native acoustic punctuation built into Whisper and Parakeet models, eliminating punctuation duplication and word trimming artifacts.
• Overlay Timer & Status Customization: Added the overlay_show_timer toggle in src-tauri/src/settings.rs and src/main.js, allowing users to hide duration counters and status labels from the floating overlay capsule while keeping the audio visualizer and critical error alerts intact.
• Full 9-Language Localization Parity: Synchronized and filled all translation keys across all 9 supported locales (ru, en, de, es, fr, it, zh, pt, tr) in src/main.js, strictly validated by automated tests in tests/frontend-static.test.mjs.
• Automated CI/CD Release Publishing: Updated .github/workflows/release.yml (releaseDraft: false) and configured osv-scanner.toml dependency filters to ensure production installer artifacts (.exe, .msi, .sig) build, sign, and publish automatically upon tag creation.

v1.0.7

What's Changed:

• Windowed-Sinc Anti-Aliasing Filter: Added a 64-tap windowed-sinc (Blackman) low-pass filter to the audio resampler in src-tauri/src/audio_recorder.rs. This filters frequencies above 8kHz before downsampling, eliminating high-frequency aliasing artifacts on sibilants (like "s", "sh", "ch") and improving local Whisper accuracy.
• Greedy Whisper Decoding: Switched the local Whisper runner to greedy decoding (--best-of 1, --beam-size -1) in src-tauri/src/whisper_runner.rs. This enforces a faster search method, yielding a 20–40% transcription speed boost for real-time dictation with negligible quality trade-off.
• Context-Aware Cyrillic Replacements: Refined how "<unk>" tokens are handled in the Parakeet model in src-tauri/src/whisper_runner.rs. The replacement to "ё" now only applies when Cyrillic text is present in the transcript; otherwise, unknown tokens are stripped to avoid corruption in Latin or Chinese dictations.
• Streamed Punctuation Downloader: Replaced full-memory buffering of the punctuation model archive with a RAM-efficient chunked downloader in src-tauri/src/whisper_runner.rs. It includes automatic cleanup of stale downloads, timeout protections, and robust error propagation for failed extractions.
• Process Buffer Safety: Added a 4,000 character safety limit for local punctuation in src-tauri/src/whisper_runner.rs. Exceptionally long dictations will now skip formatting to prevent silent process crashes caused by the Windows CreateProcess command-line argument limits.
• Terminal-Safe Dictation Mode: Added a copy_context_on_start toggle under settings in src-tauri/src/settings.rs. This allows users to disable automatic selection capture (which sends Ctrl+C to copy text) when dictating into terminals where Ctrl+C would abort running shell tasks.
• Precise Error Classification: Refined the network error checking in src-tauri/src/lib.rs to distinguish account-level failures (like quota limits or invalid API keys) from transport-level unreachable networks. This prevents authorization issues from triggering a silent, unwanted local model fallback.
• Mutex Livelock Protection: Replaced unsafe mutex lock unwrapping in src-tauri/src/lib.rs with safe fallback handlers, preventing application-wide panics if a background thread poisons a shared mutex.
• Reduced Streaming Latency: Halved the initial background streaming loop startup wait time from 4s to 2s in src-tauri/src/lib.rs to provide a more responsive real-time typing feedback loop.

Full Changelog: v1.0.6...v1.0.7

v1.0.6

What's Changed:

• API Key Auto-Trimming: Added automatic leading/trailing whitespace and newline characters trimming (.trim()) to API keys. This prevents invalid authorization headers caused by accidental copy-paste trailing spaces from AI provider dashboards.
• Stable Configuration Engine: Reverted the experimental secure keyring (system credential manager) integration. It caused credentials to disappear on some Windows configurations. All keys are safely stored back in the stable local settings.json format.
• RAII Disk Space Guard: Introduced a robust DeleteOnDrop manager for model downloads. If a large download (up to 1.6 GB) fails, times out (30 seconds of inactivity), or is manually cancelled, the file handle is dropped, and the partial .tmp file is immediately deleted to save disk space.
• Strict Network Error Propagation: Fixed a silent bug where network timeout failures during download streams were ignored, leaving corrupted model files on disk instead of flagging the error.
• Thread Safety Polish: Resolved a potential thread panic on lock poisoning in the Parakeet server manager.
• Parakeet limitations documented: Updated the documentation to note that NVIDIA Parakeet v3's transducer model in sherpa-onnx strictly requires greedy_search decoding, making custom hotword biases incompatible at the engine level.

Full Changelog: v1.0.5...v1.0.6

v1.0.5

Highlights

• Lightning-fast Parakeet Engine: Added NVIDIA Parakeet (TDT 0.6b v3) as a second offline engine. It runs as a resident background server, dropping recognition latency from ~12s to under 1 second.
• Silero VAD Integration: Replaced the old energy gate with a proper neural Voice Activity Detection. This effectively eliminates "silence hallucinations" where background noise was transcribed as random text.
• Signed Auto-updater: Releases are now cryptographically signed, and Aura can automatically check for and install updates in-app.
• Whisper Large v3 Turbo: Added support for downloading the latest high-accuracy Turbo models for local dictation.
• Cloud-to-Local Fallback: If your cloud provider is unreachable (e.g., VPN blocks), Aura will automatically fall back to your downloaded local model and transcribe the audio anyway.

Fixes & Improvements

• Model downloads can now be cancelled mid-progress.
• Local Whisper now properly utilizes all CPU cores for much faster transcription.
• Fixed an issue where the Parakeet server binaries were missing from the production installer.
• Clearer network error states for 403 blocks.

v1.0.3

Full Changelog: v1.0.2...v1.0.3

v1.0.2

Full Changelog: v1.0.1...v1.0.2

v1.0.1

Aura v1.0.1 — Hotfix & Security Release

Fixes & Improvements

• Fixed critical local Whisper crash in release build: Resolved an issue where local voice dictation returned a "Model not downloaded" error. CPU-specific instruction DLLs (like ggml-cpu-alderlake.dll) are now successfully loaded by setting the sidecar working directory directly to binaries/ at startup.
• Case-insensitive environment variables handling: Added robust detection for Windows PATH casing when launching external processes.
• Added diagnostic logs: The application now saves raw transcription errors to last_transcription_error.txt in the local AppData folder for easier offline debugging.
• Security Hardening (IPC Hardening): Restructured arguments validation in Rust backend commands to prevent path traversal risks.

---

Full Changelog: https://github.com/malashkadev/aura/commits/v1.0.1

Aura v1.0.0

• Global-hotkey dictation with push-to-talk and toggle (tap-to-latch) modes; Esc cancels a recording.

• Cloud engines: Gemini, OpenAI (Whisper + GPT), Groq (Whisper + Llama) with per-provider API keys.

• Local engine: whisper.cpp sidecar, model download manager (tiny/base/small/medium) with delete option.

• AI transcript cleanup: punctuation, grammar, filler-word removal.

• Experimental live streaming mode with smart diff typing and adaptive chunk intervals.

• Transcription history (last 50 entries) with one-click copy.

• Custom dictionary hints, 9 fixed recognition languages + auto/keyboard-layout detection.

• Optional voice punctuation commands (Russian / English).

• Overlay with live waveform, recording timer, localized status/error messages and sound themes (zen / rhodes / sci-fi / classic) with volume control.

• Autostart with Windows, tray icon, single-instance guard.

• Focus guard: simulated typing/paste never lands in a window the user switched to mid-dictation.

• Windows system proxy support and network timeouts for cloud requests.