Changelog

1.0.35

New features

  • Find images by their visual content in clipboard history. Enable visual search in settings to complement text search within images. This option is off by default.
  • Choose where to store your dataset: move it or link an existing folder, including a location managed by Drive, and manage its offline availability.
  • The Marianne voice is now available to install from the model manager for local speech synthesis.

Fixes

  • Pasting to a remote Mac through Screen Sharing is more reliable.
  • Audio settings display the Magpie voice names provided by the model.
1.0.34

New

  • Variants keep the session's history of validated takes: listen to a take, reopen it to record it again, or delete it.

Add a validated take in Variants, listen to it, reopen it to record it again, then delete the take that does not fit.

Fixes

  • In validation, the selected dictation stays in place after new takes are added: no need to find it again in the list to continue.
  • The Reviewed button state stays in sync with the saved file, even if the list reload fails after the write.
  • Transcription and Apple recognition errors now appear in your interface language, even when they happen in the background.

Improvements

  • In the validation window, the Reviewed button now works both ways: a second click removes the mark and keeps the dictation selected.
  • Core ML model updates are now more discreet: a downloaded update completes its preparation at the next launch, the current revision stays usable, and while the model is being prepared, dictation shows a clear message instead of failing.
  • Permissions are now gathered in one place in Settings: granted, refused or still-to-request access can be read at a glance, with the real state of your personal folders (Downloads, Documents, Desktop).
  • Magpie audio playback starts faster: word timings are only computed when the player displays them or a voice effect needs them. Settings flag the effects that delay the first sound.
  • Transcription model handling is more robust: only one model stays in memory, a failed load restores the previous revision, and a model change requested during dictation applies after its last result.
1.0.33

New features

  • Build a stack of text, images and files from your clipboard history. Reorder the items, then paste them together or pass them to a Smart Action. The stack remains available after restarting Oriloq.

Combine three pieces of text, choose their order and paste them together into a reply to a customer.

  • Open an image from your history to explore it with zoom and select its detected text to copy it.
  • A new Smart Action inserts today's date at the cursor, using your language's date format.
  • A new Smart Action copies the paths of files and folders selected in Finder, one path per line. It can also retrieve the path of the active document in compatible applications.

Select two files in Finder, say “Maestro, copy the paths”, then paste their paths into VS Code with Cmd + V.

Retrieve the path of the file already open in VS Code with “Maestro, copy the paths”, then paste it into Codex with Cmd + V.

Fixes

  • A brief press of the Push-to-Talk key now restores the keystroke before the next character, preserving the correct text order.
  • Magpie speech synthesis now correctly finds compatible CPU and Neural Engine models, including from chat and Smart Actions.

Improvements

  • From your history or stack, find Smart Actions compatible with your content using a search field ready to type in. Pin your favorites or hide those you don't use.
  • Smart Action voice commands better account for the request expressed in the first sentence, keeping later details from diverting the choice of action.
  • The clipboard panel features a refined layout and a centered action dock.
1.0.32

New features

  • New built-in Smart Actions let you clean up dictation, correct or rephrase text, summarize or translate it, record a meeting, transcribe a file, and capture on-screen content. Their names and voice search are available in five languages.

Turn an abbreviated draft into a complete message with Maestro.

Capture the active window with Maestro, then paste its image into a chat with Cmd+V.

  • File transcription now supports WebM and MKV, including files with multiple audio tracks.

Drag a recording onto the Oriloq Dock icon and see each speaker’s turns in the transcription window.

  • On compatible local Parakeet or Canary models, dictation can keep only the first voice it detects.
  • Speech synthesis now includes the S27 and GLaDOS V2 effects. The Marion voice is also available with compatible Magpie models.
  • A release notes journal lets you revisit notes for different versions directly in Oriloq.

Fixes

  • Pasting into Oriloq no longer gets stuck while retrieving app context.
  • The order of speakers is corrected in some stereo transcriptions.
  • Smart Actions created from chat are now found correctly in search.
  • Retries when generating text variants better respect the requested length.
  • A crash related to notifications for cleaning up pending dictations is fixed.

Improvements

  • Clipboard history search is faster and more responsive, with improved result ranking.
  • Audio playback continues even when the interface is temporarily busy.
  • The floating bubble opens more smoothly.
  • First-time installation and recovery after granting macOS permissions are more reliable. Oriloq prepares dictation before completing setup.
  • Choose how long to keep pending dictations before they are automatically deleted. Validated dictations are kept.
  • Transcription variants get better suggestions and generation settings aligned with chat.
  • Audio playback and recording Smart Actions let you choose the engine, Magpie model, voice, voice effect, and playback speed.
  • Starting and finalizing meeting recordings is more reliable.
  • Smart Action icons are now created with Image Playground when it is available on your Mac.

Engine availability

  • Local Voxtral and Nemotron transcription engines, and Kokoro speech synthesis, are no longer available. If a setting used one of them, choose a supported engine. Voxtral through the Mistral API remains available.
1.0.31

What's new

  • Oriloq's clipboard manager keeps a history of everything you copy, whatever app it came from. Open it with Option + Command + V, search by keyword or by meaning, then paste a text, an image or a file directly, including a file dragged from the Finder.
  • The clipboard manager integrates with Smart Actions: your automations can reuse copied content and its context (app, window, field, URL), save an image, or query the history from the chat.

Fixes

  • Selecting a quantized Parakeet model no longer makes Oriloq quit abruptly: the app now rejects that model before loading it, with a clear message and no pointless retry.
  • The "Ignore other voices" setting is no longer greyed out when the selected model is a local model detected automatically on your Mac.

Improvements

  • The recommended French model (Oriloq, Core ML) is updated to a newer fine-tuned version, for more faithful French recognition.
  • Downloading that model now installs only the files it actually needs to run: the transfer is faster and takes less disk space.
  • The clipboard manager panel is more reliable when changing selections and dragging and dropping items.
  • Text variant suggestions now better cover both short and long phrasings.
  • The floating icon fits in better with macOS, with a refined glass effect and a reflection that follows the pointer more naturally.
1.0.30

New features

  • The model catalog now clearly distinguishes the Core ML and MLX versions of the transcription models, with separate labels and version metadata, so you can pick the right variant without any ambiguity.

Fixes

  • Transcribing a screenshot no longer launches twice in the background: analysis restart is protected, and memory stays under control thanks to strict prompt-cache management.
1.0.29

What's new

  • Oriloq now supports transcription models fine-tuned with adapters: the model manager shows disk usage and model weights size separately, so you can choose with full knowledge.

Fixes

  • The automatic update system has been fixed: security validation is more reliable, and a clear message offers to retry or send a diagnostic to support when macOS refuses.
  • Deleting a shared model can no longer cause its base to be lost: Oriloq checks the whole local catalog, including models sharing a common base and training outputs, before any deletion.
1.0.27

What's New

  • Oriloq now tells speakers apart in transcribed files and meetings, with a bundled diarization model and refinement of every speaker turn. Speakers can be renamed after transcription, and dictation can keep only the first voice heard.

Fixes

  • Echo cancellation was upgraded to the current WebRTC version, with audio buffering that better absorbs microphone dropouts and pace differences between the capture and the reference.

Improvements

  • A real echo cancellation test is available in Settings: it keeps the microphone, the system reference and the post-cancellation output locally, with nothing uploaded automatically.
  • Meetings record system audio by default, with the microphone still offered for physical rooms.
  • Starting a recording is more robust: a repeated press no longer starts a second capture, and a held shortcut reuses the startup already in progress.
1.0.26

What's New

  • In dataset curation, mark the entries that make up your validation set and filter the list to show only those.

Fixes

  • Your custom text correction rules are now recovered during settings migration: nothing is lost when moving to a new version.
  • The misleading voice isolation warning has been removed: Oriloq no longer reports a state it cannot guarantee.

Improvements

  • Audio capture is more responsive and stable: the hardware buffer now adapts automatically to your device, with latency capped at 80 ms and recording given priority.
  • Smart Actions using a custom model provider are more reliable: model and variable validation is centralized, with clearer error messages.
  • The TurboQuant compression option has been removed: it slightly slowed generation and was rarely used. Your existing settings remain compatible.
1.0.25

New features

  • Automatically discover and add language-model providers available on your Mac from the Model Manager.

Fixes

  • Segmentation settings no longer show two Parakeet-specific warnings that did not apply to your configuration.

Improvements

  • Update checks now begin at launch, so a new version is reported sooner.
  • For new settings, dictation now waits three seconds of silence before stopping. Your existing custom settings are preserved.
1.0.24

What's New

  • Set up without starting over: interrupted model downloads can resume, models already on your Mac are preserved, and migration and permission steps are better guided. The DMG uninstaller also lets you choose whether to keep your data or remove it completely.
  • One model picker for chat and settings: pin favourites, see each model’s image, audio and reasoning capabilities, then open its advanced settings.
  • Turn chat history into a workspace: import, export, rename, duplicate or delete one or several conversations, including in batches.
  • Send selected text to Oriloq from the macOS Services menu, and reuse a selection made in chat in the floating icon’s context.
  • Stay in control when the same Smart Action is launched more than once: a new dedicated step lets you wait for earlier runs to finish, interrupt earlier runs and cancel the new one, or interrupt earlier runs and continue with the new run.

Fixes

  • Dictation pasting is more reliable and preserves your clipboard more consistently, including when the destination app changes at the last moment.
  • Continued dictation preserves word boundaries more naturally and correctly replaces punctuation at the end of a field.
  • VAD timing markers, silences and reconstruction thresholds are better aligned with the segments that were actually transcribed, so audio saved for validation now reflects those segments.
  • After echo suppression, Oriloq no longer unexpectedly falls back to raw audio.
  • Your manually selected microphone stays reliably associated with its device, even after its connection or device order changes.
  • Canary timestamps are safely finalised before a recording is validated, keeping word timing consistent with its transcription.
  • Conversations no longer replay old tool calls, retry empty agent responses, and recognise Ollama provider capabilities more accurately.
  • Selected-text capture works better in xterm.
  • In curation, when a response already contains a sample preview, the list summary no longer overwrites it. Long inspections now return clearly marked excerpts when truncated, keeping the session readable.

Improvements

  • Text-variant suggestions display their diagnostic, also check for varied sentence lengths, and can automatically retry when they remain incomplete.
  • Recording validation displays the transcription method and indicates whether text comes from pure ASR.
  • Custom model providers are described more clearly in chat and settings, including their models and available capabilities.
  • Chat applies the selected evaluator provider from the start and isolates the TurboQuant prefix cache between conversations.
  • Smart Actions generated by the agent now favour the shortest useful workflow, ask for clarification when information is ambiguous, and refuse to paste an invented value.
  • Semantic search now takes Smart Action names and descriptions into account.
  • Adjust speech speed directly from the OmniVoice HUD, with a cleaner playback interface.
  • After an update, compatible panel widths are preserved while older incompatible values are cleanly reset.
1.0.23

What's New

  • New audio conversion Smart Action: convert your audio files between formats, with an option to replace the source file on the fly.
  • Configurable speech and chat model selection: choose exactly which model to use for voice and conversations, straight from the settings.

Fixes

  • Audio playback resumes correctly after an audio device change (Bluetooth headset, USB interface).
  • TTS pause state is preserved when a streaming session starts late.
  • The push-to-talk and recording activity lifecycle is more reliable.
  • Fixed a crash when opening the dataset validation window.
  • Bluetooth dictation audio availability is more reliable.

Improvements

  • Seamless Bluetooth audio and lossless FLAC recordings: device changes are handled instantly and voice playback resumes from the exact frame where it stopped, with no audible rewind.
  • OmniVoice speech synthesis is unified and playback speed is now remembered across sessions.
  • Core ML runtime selection for OmniVoice, to optimize performance based on your Mac.
  • Clearer model memory compatibility indicators, so you can tell at a glance whether a model can run on your machine.
  • Embedding engine migrated to Qwen3 DWQ, with warnings for partial semantic index and automatic retry of failed migrations.
  • Support for structured Gemma 4 tool results for more precise Smart Actions.
  • More responsive interface: model scanning, symbol search, and catalog management now run in the background without blocking the UI.
  • Optimized transcription highlighting: a single audio clock drives the active segment highlight, reducing system load.
  • Proactive release of Smart Action audio resources after playback.
  • The Smart Action editor saves more efficiently while typing, without stutter.
  • Improved transcription pipeline and UI state handling for more stable sessions.
1.0.22

What's New

  • New pure ASR transcription model: Nemotron performs true streaming during recording, so text appears as you dictate.
  • New configurable neural voice: describe a voice or clone your own to read text aloud, record podcasts, listen to selected on-screen text, and enhance smart actions.
  • Chat and smart actions can now answer aloud.
1.0.21

New Features

  • Parakeet dataset curation and fine-tuning: prepare, validate, and improve transcription datasets directly in Oriloq, with a dedicated workflow to train a model that better fits your usage.
  • Faster Smart Action composition: step search makes profile creation easier and helps you find the right action block without browsing the full list.
  • Gemma 4 QAT models: Oriloq replaces the previous MXFP4 models with the new Gemma 4 QAT models for more reliable local conversations.

Fixes

  • More predictable GPU transcriptions: intensive ASR transcriptions are better coordinated to avoid concurrent processing that could disturb results.
  • More reliable scrolling capture: long captures better avoid duplicates and end-of-page offsets.
  • More accurate text corrections: numbers and some replacement rules are recognized more reliably, reducing unexpected corrections.
  • Better protected transcription data: incomplete segments are rejected earlier to avoid feeding datasets with partial content.

Improvements

  • Smoother AI chat: Gemma streaming and conversation display are more stable and responsive.
  • Clearer tool results: outputs produced by Claude and Gemma are better structured, making it easier to understand what was executed.
  • Stronger audio flow: recording, segmentation, and the ASR lifecycle are more stable during long sessions.
  • Clearer Parakeet fine-tuning: base model configuration and GPU tracking are better guided during training.
1.0.19

New Features

  • Vocal effects in chat: chat voice synthesis now includes sound effects (GladosKit) for a livelier, more personalized experience.
  • Text correction diagnostics: a new tool lets you analyze and understand the corrections applied to your transcriptions, giving you greater control over the final result.

Fixes

  • More reliable chat: native Markdown rendering eliminates deadlocks and stabilizes bubble height during generation.
  • Gemma inference priority respected: priority parameters and parallel generation defaults are now properly honored.
  • Update restart fixed: the application restarts correctly after installing an update.
  • Smoother floating icon: floating icon transitions are more natural and short-dictation latency is reduced.
  • Model labels fixed: model names display correctly in all languages.

Improvements

  • Audio dataset more reliable: audio capture and microphone latency are better managed for more stable recordings.
  • Optimized configuration propagated: the optimized model configuration automatically applies to associated Smart Actions.
  • Optimizer display improved: the AI optimizer interface is clearer and more responsive.
1.0.18

New

  • Multi-model AI optimizer: create and compare prompt variants with Apple Intelligence, Gemma, or Ollama, with live tracking of scores, tested examples, and the best prompt.
  • License activation through Oriloq links: an activation link now opens the License tab directly, starts activation, and clearly shows the status or error without manually copying a key.

Fixes

  • Better contextual punctuation preservation: dictation and terminal prompts now preserve punctuation around replaced text more reliably.
  • More reliable Smart Actions with Ollama: reasoning options and execution timeouts are handled more consistently to avoid incoherent results or stalls.
  • More useful Smart Action notifications: clicking a notification now opens the editor on the related action.
  • More stable floating icon: the overlay stays visible without stealing focus and closes more cleanly when it has nothing left to show.
  • More reliable Oriloq menu: Quit and chat access from the menu bar now behave more consistently.

Improvements

  • More stable local chat and parallel generations: generation sessions reuse their context more reliably and isolate model loading to reduce collisions.
  • Refined model recommendations: transcription model suggestions better account for the app language.
  • Simplified Ollama settings: enabling Ollama, testing the connection, and selecting models are clearer in settings and Smart Actions.
  • Stronger model diagnostics: Oriloq keeps more context about sessions and tool results, making complex behavior easier to analyze.
1.0.17

What's new

  • TurboQuant control exposed in the chat: a new setting lets you enable or disable TurboQuant directly from the Gemma chat inference parameters, without diving into advanced settings
  • Per-model TurboQuant guidance: a dedicated tooltip details the real impact on speed and memory for each Gemma model (E2B, E4B, 26B-A4B, 31B), with a clear recommendation based on your context size

Improvements

  • More reliable Gemma model switching: quickly chaining several model selections can no longer leave the chat in an inconsistent state; the previous engine is unloaded cleanly before the new one is loaded
  • Prefix KV cache synced with TurboQuant: the prefix KV cache is automatically disabled when TurboQuant is off, preventing incompatible combinations
1.0.16

What's New

  • Gemma download during setup: the local chat model starts downloading right from the first launch, so you can access local chat faster after onboarding
  • Improved onboarding: required installation steps are now presented more clearly, for a smoother getting-started experience

Fixes

  • Auto-gain and speech detection: input auto-gain now only activates when speech is detected, preventing unwanted amplification during silences
  • Translation cold start: translation no longer freezes on the very first startup after installation
  • Dictation after dataset correction cancel: dictation now resumes correctly after cancelling a correction in the dataset module
  • Chat streaming height: markdown rendering no longer causes height jumps during text generation
  • Window cleanup: resources from closed windows are now released properly, improving overall stability
1.0.15

New Features

  • Enriched Dictation Smart Action : a new smart action that enriches your dictation with whatever you copy to the clipboard while you speak. A screenshot is described and woven into your text ; copied text is inserted cleanly, as if you had carefully woven it into your narrative, all without interruption.

Fixes

  • Reliable Smart Action Cancellation : cancelling an ongoing smart action now works consistently, even while the model is loading
  • Robust Scrolling Capture : scrolling-area capture has been hardened; reconstructed overlays and segment stitching are now more reliable
  • Local Chat Stability : in case of a generation error, the Gemma engine shuts down properly and releases the GPU, preventing hangs
1.0.14

New Features

  • Smart Action Agent: a new intelligent agent automatically discovers available capabilities and executes your smart actions hierarchically, for smoother automation
  • Ollama Chat: integration of chat via Ollama, allowing you to converse with local models served by your Ollama instance
  • Duplicate Conversation: in the Gemma chat, you can now duplicate an existing conversation to create an independent copy
  • Intelligent Chat Memory Management: the chat model automatically unloads after 5 minutes of inactivity, freeing RAM without any action on your part

Fixes

  • Robust User Messages: fixed a case where an optional user message could cause unexpected behavior in chat

Improvements

  • Smart Actions Default to Gemma 4: new smart actions now use the Gemma 4 engine by default for better transformation quality
  • Chat Scroll and Auto-Loading: automatic scrolling in chat is smoother, and conversations reload correctly when editing
  • Improved Chat Audio Workflows: audio file handling in chat is more fluid and predictable
1.0.13

What's new

  • Next-generation multimodal local chat: Oriloq now ships a 100 % local chat built on the Gemma family of models, able to understand text, images and audio. Backed by a state-of-the-art implementation (TurboQuant quantization, MLX inference, very large contexts) that runs on a local machine with minimal RAM, it is very likely the best implementation available today on macOS
  • Smart Actions powered by local chat: smart actions can now rely on this local engine to transform your text without any network call
  • Local Claude Code server: Oriloq exposes a server compatible with the Claude Code CLI right on your machine, so you can use your local model as a reasoning backend without any external dependency
  • Smart Actions import / export: your smart actions can now be shared through a portable JSON format, with automatic handling of relative file paths
  • Dataset spell check: fine-tuning dataset validation now includes built-in orthography correction

Improvements

  • Much faster dataset: loading a dataset entry is now nearly instant thanks to caching of audio duration and file hashes
  • Better organized Voxtral settings: the Voxtral language choice has been moved into the transcription settings, closer to related options
  • Sturdier capture & recording: audio capture and recording flows have been hardened to better handle edge cases

Fixes

  • Cleaner relaunch: a lingering Oriloq instance can no longer survive a new build and mask subsequent errors
1.0.12

Fixes

  • Microphone reliability: fixed an issue where the "Microphone is not responding" error appeared repeatedly, including with the Mac's built-in microphone
  • Audio gaps eliminated: word dropouts during dictation caused by system load contention have been fixed

Improvements

  • Faster dictation start: dictation now starts more quickly by reusing the audio engine prepared in advance
  • Complete dictation capture: the last words spoken just before stopping dictation are now fully captured, even over Bluetooth
1.0.11

What's New

  • Smart screenshot capture: the three capture modes (full screen, selected area, active window) are now unified in a single menu, with a new active window mode that automatically captures the frontmost application
  • Model memory management: a new option in settings lets you keep transcription models loaded in memory between dictations for near-instant startup; unloading adapts automatically based on your Mac's memory pressure
  • Voxtral streaming (beta): the first Voxtral streaming mode settings are available, paving the way for real-time transcription as you speak

Bug Fixes

  • Multi-channel file transcription: audio compression now uses a unified speech map, fixing potential artifacts when transcribing stereo or multi-track files
  • Simplified Bluetooth capture: removed a residual dual audio capture in Bluetooth mode that could consume resources unnecessarily
1.0.10

What's New

  • Echo Cancellation (AEC3): Oriloq now automatically filters out system sounds during dictation using WebRTC AEC3 technology, delivering cleaner transcriptions even when your Mac is playing audio
  • Meeting Recording: echo cancellation is also available for meeting recordings, preventing participants' audio from interfering with microphone capture
  • Wired Microphone Safety Net: when your Bluetooth headset has connection latency, Oriloq temporarily switches to the built-in microphone so you never miss a word, then smoothly transitions back to Bluetooth
  • Meeting Smart Action: a new smart action dedicated to meeting recording has been added

Fixes

  • Volume Restoration: system volume is now reliably restored after each dictation session, for both Bluetooth and wired setups
  • Smooth Audio Transition: switching between the backup microphone and Bluetooth now uses a 300 ms crossfade to avoid audible cuts
  • Phonetic Correction: improved French phonetic pronunciation rules for more accurate transcriptions
  • Model Stability: CoreML models stay loaded in memory and MLX idle unloading is more reliable

Improvements

  • AEC Disabled by Default in Bluetooth Meetings: echo cancellation is off by default in meeting mode when using a Bluetooth headset, to avoid unintended audio alteration; can be enabled on demand
1.0.8

New Features

  • Multi-model comparison: compare results from multiple transcription engines side by side when validating your datasets, to choose the best model for your needs
  • Code editor compatibility: Oriloq automatically detects Electron-based editors (VS Code, Cursor) and displays a notification to guide you through enabling accessibility

Bug Fixes

  • Free tier user license: fixed a false security alert that could appear for users not using smart actions
  • Smart action naming: references to the transcription engine in smart actions now correctly display "Apple Intelligence" and "AI" instead of internal technical names
  • Text corrections toggle: the text corrections toggle button now works more reliably
1.0.7

New Features

  • Word-level timestamps (ElevenLabs): ElevenLabs cloud transcription now extracts timestamps at the word level, with automatic punctuation correction and precise alignment

Bug Fixes

  • Faster pasting: significant latency reduction when pasting segmented transcriptions, thanks to per-segment phonetic corrections and parallelized processing
  • Multitouch stability: fixed a crash related to multitouch trackpad during sleep/wake cycles
  • Menu bar icon: the Oriloq icon now displays correctly from the moment the application starts
  • Dataset validation: the "What the model understood" section is now always visible during dataset validation

Improvements

  • More reliable scrolling capture: better uniform background detection, improved bottom-of-page content handling, and optimized junction overlap
1.0.6

What's New

  • Cloud transcription for everyone: ElevenLabs Scribe and Mistral Voxtral engines are now available directly in the model selector — no developer mode required
  • Scrolling screenshot capture: Capture an entire page with automatic scrolling and pixel-perfect stitching — perfect for long documents and web pages
  • Smart Action Mode: A dedicated shortcut to trigger your smart actions, with a Beta badge and try-it area in the onboarding screen
  • Word deletion rules: Text corrections now support automatic removal of unwanted words or phrases
  • Redesigned dataset validation: The dataset validation interface has been completely simplified for clarity

Bug Fixes

  • Text corrections: Typing a word to delete no longer causes the text field to lose focus
  • Responsive push-to-talk: Rapid space bar presses no longer drop characters in push-to-talk mode
  • Reliable cloud transcription: Cloud results now return consistent plain text, with the option to re-transcribe locally
  • Punctuation-aware corrections: Replacement rules containing punctuation now apply correctly
  • Stable Bluetooth: Switching Bluetooth devices no longer interrupts an active recording
  • Optimized push-to-talk: Default delay set to 250 ms for a smoother dictation experience

Improvements

  • Automatic memory release: ML models automatically unload from memory after a period of inactivity, reducing the memory footprint
  • Streamlined interface: Chat mode removed to refocus the experience on transcription
  • Unified settings: Settings sections have been consolidated for smoother navigation
  • Memory diagnostics: A real-time memory usage indicator is available in developer settings
1.0.5

New Features

  • Audio & video file transcription: drag and drop an audio or video file directly onto Oriloq to transcribe it — no manual input needed, everything works via drag-and-drop
  • Built-in audio player: follow the transcription in real time with a player that highlights segments as they play
  • Multi-speaker display: file transcriptions are shown in conversation mode with dedicated chat bubbles for each speaker
  • Flexible export: export your transcriptions as JSON, SRT (subtitles), or copy them straight to your clipboard
  • On-the-fly model selection: switch transcription models to re-transcribe a file with greater accuracy
  • Automatic audio channel detection: Oriloq identifies separate audio streams for better speaker attribution
  • Adaptive auto-gain: microphone volume adjusts automatically during dictation for optimal capture

Fixes

  • Sleep/wake stability: fixed a crash that could occur when putting the Mac to sleep or waking it up
  • More reliable pasting: eliminated a race condition that could occasionally prevent transcription from being pasted

Improvements

  • Segmentation guide: a new educational popup now explains how automatic segmentation works and which models support it
  • Silence compression: silent passages are automatically compressed during file transcription for faster results
  • Update cleanup: automatic detection and removal of pending updates to free up disk space
1.0.4

Bug Fixes

  • Reliable updates: fixed a critical issue that could close the application without relaunching it during an automatic update
  • Stable relaunch: the restart mechanism after an update is now more robust, ensuring the application reopens properly

Improvements

  • Controlled updates: updates no longer install automatically at startup — you choose when to restart via the dedicated button
  • Failure recovery: if an update installation fails, the application automatically restores the previous version
1.0.3

What's New

  • Launch at Login: Oriloq can now start automatically when you log in to macOS
  • First-Launch Tutorial: an interactive guide walks new users through the setup step by step
  • Independent Segmentation: disable audio segmentation while keeping voice detection active, for greater flexibility

Bug Fixes

  • Model Persistence: the selected model is now properly preserved during dataset retranscription
  • Process Cleanup: concurrent Swift process cleanup is now scoped to the current workspace, without interfering with other projects

Improvements

  • Lower Segmentation Latency: async/await refactoring for more responsive audio processing
  • Streamlined Translation: simplified built-in translation service for better performance
1.0.2

What's New

  • Automatic updates: Oriloq now checks for new versions and offers one-click updates, with code signature verification to ensure authenticity
  • Redesigned model manager: new full-page interface with RAM usage statistics, size categories, and MXFP4-Mixed model support for a better quality/memory trade-off
  • Smart clipboard restoration: after each pasted transcription, your previous clipboard content (rich text, images, files) is faithfully restored
  • Automatic volume normalization: dataset audio files are automatically normalized for comfortable listening during validation
  • Smart Action icons in notifications: instantly identify which smart action produced the result

Bug Fixes

  • Fixed a memory leak when switching transcription models
  • Push-to-talk mode no longer blocks system shortcuts (Cmd+C, Cmd+V, etc.)
  • Fixed an off-by-one cursor offset in contextual adaptation
  • Improved voice activity detection (VAD) reliability
  • Fixed clipboard restoration in certain cases without Accessibility

Improvements

  • Near-instant pasting: new paste-first architecture eliminates the ~1 second delay before pasting
  • 67% memory reduction for smart action embeddings (6 GB → 2 GB)
  • Enhanced network security with hardened App Transport Security (HTTPS enforced on all domains)
  • The model picker now only shows downloaded models for better clarity
1.0.1

What's New

  • Local voice transcription: transcribe your voice directly on your Mac, without sending your data to the cloud. All inference runs on your Apple Silicon GPU thanks to MLX.
  • Automatic updates: Oriloq now updates silently in the background. A simple restart is all it takes to apply the new version.
  • Enhanced security: Apple code signature and notarization verification for every downloaded update.
  • Smart Actions: trigger intelligent actions from your transcriptions, with customizable icons, a built-in editor, and SF Symbols picker.
  • Model manager: new full-page interface to browse, download, and manage your transcription models, with MXFP4-Mixed variant support.
  • Smart picker: the model selector now only shows models present on your Mac.
  • Voice detection and filtering: the VAD engine analyzes your audio in real time to only transcribe passages containing speech.
  • CLI mode: transcribe audio files directly from the terminal with oriloq-generate.
  • Volume normalization: dataset audio files are automatically normalized for comfortable listening.
  • Multilingual: interface available in French, English, Spanish, German, and Italian.

Bug Fixes

  • Fixed a noticeable paste latency (~1 second) with a new paste-first architecture.
  • Fixed clipboard restoration for rich content (formatted text, images).
  • GPU memory leak resolved: switching models now properly releases the old model's memory.
  • Better end-of-sentence detection by the VAD (Voice Activity Detection).
  • Fixed a race condition between clearing and loading logs.
  • Fixed memory leaks in the audio pipeline and smart actions.

Improvements

  • Drastic reduction in memory usage for embeddings (6 GB → 2 GB).
  • VAD now runs on a dedicated high-priority thread for maximum responsiveness.
  • Hardened network security (strict App Transport Security, TLS 1.2+ required).
  • Strengthened licensing system with server-side version verification.
See what's coming next