What changed in each release

Version 1.8.1

No raw audio is archived. The active server database retains event transcripts for 90 days; downloaded exports and backups have separate lifecycles.

• During an Übercast broadcast, Heard text and each produced event translation are saved to the private Client Dashboard, independently of the short live-caption buffer.
• Pending passages are persisted on the Mac and retried after interruption. The app reports server acknowledgement separately from local capture and save counts.
• Clear is unavailable while broadcasting or while archive work remains pending, uncertain or failed. Clearing the Mac later does not delete the server archive.
• Client Dashboard Save Transcript downloads the complete saved record as Markdown, including all saved languages and published corrections, independently of view filters.

Version 1.8.0

Classifier requests are optional and are sent directly to TypeSafe AI only while the feature is enabled. Availability and charges depend on your TypeSafe account.

• Added an optional Classifier sidebar in both Übersetz and Übercast. The operator can switch back to Advisor or Questions without changing the transcript.
• After explicit activation with a user-provided TypeSafe API key, Jev assesses recent Heard text and paired text from the selected translation. The key is stored in macOS Keychain.
• Shows Clarity, Relevance, Reasoning support and Actionability for Heard text, plus Translation accuracy and Translation quality for the selected target language.
• Scores are indicative and do not verify facts. Translation accuracy compares displayed text, not original audio; low-confidence results are marked.

Version 1.7.2

This release skips version 1.7.1. Test the complete Übercast workflow on the venue network before an event.

• Übercast remembers the last submitted activation code until the Mac app closes, including after a failed activation attempt. The code is not saved to disk.
• Clear is disabled during an active listening session in Übersetz and Übercast, preventing accidental loss of the live Heard and Translation text.
• The direct-download Mac app checks for signed updates quietly at every launch when automatic checks are enabled. Sparkle's regular scheduled checks continue.
• The Übercast attendee player retains earlier captions across a broadcast restart and marks the start of the new broadcast.

Version 1.7.0

Questions require a server-assisted Übercast event and are private to the operator's Mac. Test the complete event workflow on the venue network before relying on it.

• Added Dual View for Heard and Translation side by side, synchronized manual scrolling, and a window-width layout with a fixed-width Advisor or Questions panel.
• Added private attendee questions to server-assisted Übercast. The operator reviews and submits them into the Mac's Heard and Translation transcripts; attendee pages do not display the questions.
• Translated questions separately into each permitted event target language and allowed the operator to switch the displayed target while listening without changing attendee feeds.
• Rebuilt local speaker attribution and improved long-session transcript retention. Speaker labels remain estimates rather than verified identities.
• Improved attendee caption scrolling and language controls. Transcript panes follow new text at the bottom and preserve manual scrollback.

Version 1.6.3

Gemini features require your own compatible Google Gemini API key and remain subject to Google's model availability, quota, pricing and regional access.

• Upgraded Gemini Translate's Advisor to Gemini 3.8 Live with one combined text-and-audio response.
• Streams the Advisor response text and native Gemini voice together, with replay through the selected output device.
• Gemini automatic summaries now review new translated text every 10 seconds; GPT remains on its one-minute cadence.
• The Advisor follows new output while at the bottom, pauses during manual scrollback, resumes at the bottom and respects Reduce Motion.
• The Advisor prompt reflects provider capabilities: Question or URL for GPT and Gemini, and Ask a question for other models.

Version 1.6.2

Test the complete broadcast on the venue network before an event. Network restrictions may affect audio connectivity.

• Added server-assisted Übercast delivery with event destinations and permitted concurrent target languages configured by the organizer.
• Added optional organizer-provided Gemini credentials for an event.
• Improved broadcast caption buffering and direct-connection recovery.
• The overlay follows the selected Heard or Translation tab and can be resized without scaling the text.
• Smoother transcript and overlay scrolling preserves manual scrollback and respects Reduce Motion.

Version 1.6.0

Übercast organizer access is arranged separately. It uses direct WebRTC without a media relay, so network compatibility and the broadcaster’s capacity must be tested before an event. The Mac must remain awake and running Übersetz.

• Added Übercast: share live translations and translated speech directly from your Mac to attendee browsers, with no attendee app required.
• Activate using an organizer code, then share a session link or QR code. Copy the URL or save the QR code as a printable PNG.
• Attendees can save or share their current translated text locally as Markdown.
• Gemini interpreting and broadcasting use its native generated translation voice.
• Updated Sparkle shutdown handling to support relaunch after installing an update.

Version 1.5.0

Version 1.5.0 can be installed automatically by version 1.4.2 through Sparkle. Übersetz uses a signed delta update when available and falls back to the complete signed and notarized DMG when necessary.

• Added an optional Interpreter for microphone conversations, speaking completed translations through the selected audio output.
• Uses GPT and Gemini provider voices for their cloud modes, while Apple Translate, Whisper-based, local, and custom modes use on-device speech.
• Keeps Interpreter unavailable with System Audio and adds headphone guidance to reduce the risk of translated speech feeding back into the microphone.
• Improved Apple Translate transcript flow so text remains connected and starts a new paragraph only after a longer pause and a completed sentence.
• Hardened Apple Translate against stalled translation requests and microphone capture interruptions.
• Moved Save into the top toolbar and aligned the translation controls consistently across providers.

Version 1.4.2

Version 1.4.2 is the first public Sparkle-enabled release. Existing installations through version 1.4.1 cannot discover it automatically and must install this DMG manually once. After version 1.4.2 is installed, subsequent signed updates can be downloaded and installed in the app.

• Added signed automatic updates for the direct-download macOS app, with a scheduled check every six hours and an immediate Check for Updates command in the application menu.
• Downloads and verifies available updates in the background, then shows a compact Restart to update message when installation is ready.
• Stops an active translation session cleanly before installing and relaunching the app.
• Added signed Sparkle delta-update support with a full-download fallback for future releases.
• Hardened app termination during an updater-requested relaunch.

Version 1.4.1

Automatic summaries use the same regular Advisor provider and credentials as manual analysis. The feature remains available only while GPT or Gemini is selected; Apple Translate remains unaffected.

• Added an optional Automatic summaries toggle for GPT and Gemini. In version 1.6.3, Gemini reviews new translated text every 10 seconds while GPT keeps the original 60-second cadence.
• Removed the separate silent observer connection, preventing a second Realtime or Live stream from running alongside translation.
• Kept automatic analysis fully inactive when its toggle is off and limited each scheduled summary to text added since the previous successful invocation.
• Moved the Automatic summaries control into the Advisor header and reduced the height of the surrounding controls.
• Improved Advisor instructions so summaries omit commentary about cut-off transcript fragments and focus on completed substantive content.

Version 1.4.0

Apple Translate remains fully on-device and does not use an audio-rescue fallback. Advisor analysis and speech remain unavailable with Apple Translate itself.

• Made normal Apple Translate use a lower-latency speech cadence, while High Fidelity retains longer context for quality.
• Kept Apple Heard and Translation as flowing text, starting a new paragraph only after at least a two-second pause and a completed sentence, and repaired retained mid-sentence chunk breaks.
• Preserved regional and script-specific locale identifiers and tightened Whisper filtering to reject unreliable short or noisy recognition.
• Removed Uncertainty Rescue audio retention and controls from both Apple Translate modes.
• Restored supported Advisor speech to a natural female voice at normal conversational volume across OpenAI, Gemini, local and custom providers.

Version 1.3.2

This release changes Gemini's live audio transport only. Speaker labels remain local estimates, and Google model access, quota, billing, regional availability and Preview-model changes still apply.

• Fixed Gemini Live Translate sessions that could stop after a few sentences, especially when speaker diarization was enabled, by keeping the selected live audio streaming continuously to Google.
• Kept local speaker attribution independent from Gemini provider turns, so diarization can label the retained transcript without pausing the cloud audio stream.
• Replaced the manual Gemini speaker-activity gate introduced in version 1.2.1 with the provider's continuous-streaming and automatic speech-detection behavior.

Version 1.3.1

The new-install default does not overwrite an existing user’s saved translation-model choice. Uncertainty Rescue remains bounded by the rolling in-memory audio window described in the privacy documentation.

• Added a Welcome guide on first launch and a What’s New summary after each app update. Both can open the translation-model picker, and the guide can be reopened from Help → Welcome & What’s New.
• Made Apple Translate the default for new installations and added a Pick a different Translation Model action to the Ready view.
• Improved GPT and Gemini stream assembly so Heard text preserves complete chunks and word boundaries instead of occasionally collapsing to a few words or joining words together.
• Kept the Heard tab visible for one second after its first text appears before automatically showing a received GPT or Gemini translation.
• Removed Uncertainty Rescue from GPT and Gemini modes, and kept transcript text stationary when rescue controls appear on hover in supported modes.
• Made Heard text editable during Uncertainty Rescue, with a fresh translation preview before applying the correction and an Undo action afterward.
• Serialized Apple Translate requests and hardened transcript-range handling so corrections cannot overlap native translation work, leave the correction toast stuck, or stop the listening session with an invalid text range.

Version 1.3.0

Uncertainty Rescue uses a rolling audio buffer of up to 90 seconds held in memory during the live session. It is not an audio archive, is not included in saved Markdown sessions, and older or restored lines cannot be replayed after their audio leaves that window.

Confidence markings are evidence from local Whisper, not a guarantee that unmarked words are correct. Cloud and other provider modes can offer review for recent lines, but generally do not expose comparable word-level confidence values.

• Added an app-wide Day appearance with a satin ivory palette, dark-gray type and a matching translucent overlay. The saved Day or Night choice remains tonally consistent through normal and overlay transitions.
• Replaced the main text-labelled actions with compact icon controls, including the sweeping-broom Clear action.
• Simplified overlay catch-up into a standard scrollable transcript: it follows new speech while at the bottom and pauses automatic scrolling when you move up.
• Added File menu commands to open a saved Markdown session, save the current non-empty session, clear it, or reopen a recent Markdown file.
• Added Uncertainty Rescue for recent completed lines. Measured low-confidence words from local Whisper receive a dotted amber underline; recent lines with buffered audio can still be reviewed when a provider supplies no confidence score.
• A rescue pass re-listens with the on-device Whisper model, displays the current and proposed source text and translation, and applies both together with Undo. The corrected text is translated by the currently selected provider without web search.

Version 1.2.1

Speaker labels remain estimates rather than verified identity. Accuracy still depends on the voices, audio quality, overlap, selected provider and chunk boundaries; review important attributions against the conversation.

• Reworked speaker diarization into modular turn tracking with stable audio-interval IDs, helping delayed transcription and translation results remain attached to their originating turns.
• Improved local translation alignment with ordered transcription and translation queues, timestamp-aware fragments, explicit speech-split reasons and bounded context carry-over.
• Serialized Gemini speaker activities so a new locally detected speaker waits for the preceding Gemini provider turn to complete instead of sending overlapping turn signals.
• Made Clear reset the retained session, speaker numbering, detected-person count and learned diarization state.
• Fixed intermittent System Audio restart failures after Stop with generation-safe teardown and bounded retry handling.
• Restored Escape from Overlay to Normal mode even when the overlay has not been clicked or focused.
• Increased the macOS Liquid Glass blur and refraction, and added a Speaker diarization Beta guide explaining model differences, failure modes and deliberate-use safeguards.

Version 1.2.0

• Added Gemini Live Translate as an opt-in Preview provider with an independent, user-owned Google Gemini API key stored in macOS Keychain.
• Added Gemini-matched Advisor analysis, selected-passage Google Search grounding and Gemini Advisor speech. Google pricing, quota, project, regional and Preview-model availability apply.
• Reorganized Settings into GPT Realtime, Gemini Live, Apple Translate, Local models and Custom endpoint provider families.
• Made the Local models picker display one row per model family, prefer an equivalent installed Ollama copy, and preserve the four managed Qwen downloads without requiring Ollama.
• Added a versioned Add local model catalogue for curated official Ollama Library tags, streamed pull progress, automatic refresh and selection, plus an advanced untested-tag field.
• Clarified that model-catalogue inclusion is a runtime compatibility review rather than a translation-quality benchmark.
• Fixed Apple Translate target-language preservation and cold-start local-model selector races.
• Improved Gemini setup timeout and provider-error reporting without silently switching providers.

Version 1.1.6

• Fixed mode-switching edge cases so interrupted transitions finish coherently without leaving the window hidden, transparent or at the wrong saved frame.
• Added a more legible blue Liquid Glass background and a native flowing morph between normal and overlay modes on macOS 26, with vibrancy and Reduce Motion fallbacks.

Version 1.1.5

• Changed normal Advisor analysis from a single contextual insight into a concise conversation briefing covering the central situation, important facts and claims, apparent intent, decisions, next steps and unresolved questions.
• Kept selected-text analysis as a shorter focused explanation and reserved OpenAI web search for selected passages where external context is materially useful.
• Increased local and custom Advisor output allowance and added one bounded retry for token-limited or visibly unfinished responses.
• Prevented incomplete Advisor output from being saved or spoken as a finished result, and disclosed when earlier transcript text was omitted to fit the model context window.

Version 1.1.4

• Limited GPT Realtime overlay captions to the latest four rendered lines so the overlay no longer grows indefinitely during continuous translation.
• Added a gentle 0.2-second upward scroll when a newly wrapped line appears at the bottom of the overlay.
• Preserved the complete GPT translation in the session transcript while bounding only the visible overlay viewport.
• Kept non-GPT adaptive overlay sizing and Plotter mode behavior unchanged.

Version 1.1.3

• Added optional Plotter mode for non-GPT Realtime providers, revealing completed Heard and Translation chunks word by word in the normal and overlay views.
• Normal and overlay window positions and sizes are now remembered independently and restored when switching modes or reopening Übersetz.
• Right-to-left Heard and Translation text now uses RTL direction and right alignment; Plotter overlay captions align left for LTR languages and right for RTL languages.
• Kept the main translation controls on one horizontally scrollable row and reduced the target-language selector width.

Version 1.1.2

• Removed blank lines from overlay captions and made the overlay expand automatically to show the complete current caption.
• The overlay returns to its compact default height when the next caption needs less space.
• Added Check for Updates to the Übersetz application menu with current-version, download and connection-error feedback.
• Enabled the macOS Settings menu command and made Settings and update checks return from overlay mode to the normal window automatically.

Version 1.1.1

• Added a per-endpoint Translation cadence setting with Balanced and Low latency modes.
• Low latency translates shorter speech segments more often and sends very short fragments without waiting for the next phrase.
• Translation cadence works with Ollama and OpenAI-compatible endpoints and is saved separately for every endpoint.
• Improved GPT-5.6 compatibility on OpenAI-compatible endpoints by using max_completion_tokens and the model's supported default temperature.

Version 1.1.0

• Added Apple Local translation using Whisper language detection and Apple's on-device translation framework.
• Apple Local supports dynamic target languages and defaults to Low Latency mode, with High Fidelity available where supported.
• Apple Local requires macOS 26.4 or later. Other Übersetz providers continue to support macOS 15+.
• The Heard tab now shows the detected source language.
• Selected transcript text can now be sent directly to the Advisor for explanation.
• Added Advisor voice-speed controls with 1.0×, 1.25× and 1.5× playback.
• Cloud translation is prevented from starting while the Mac is offline.
• Fixed a crash affecting GPT Advisor audio playback.

Version 1.0.9

The public DMG is Developer ID signed, accepted by Apple's notarization service, stapled and accepted by Gatekeeper. See Download verification for the current filename, size, checksum and local verification commands.

• Translate microphone input, mixed macOS system audio, or both.
• Use GPT Realtime with your own API key, built-in Whisper and Qwen combinations, or a compatible custom endpoint.
• Follow separate Heard and Translation views with newest-text emphasis.
• Use the movable two-line overlay and contextual Advisor explanations.
• Attribute speakers, clear the retained local session and export completed translations as Markdown.
• Download and verify optional local models with resumable transfers and SHA-256 checks.