What's new in SideSense Pro

Updates no longer freeze on machines where antivirus interferes with the updater.

• FixedUpdating could freeze the PC on machines where antivirus software (seen with Norton) blocked part of the update: the updater's are-we-still-running check could end up counting itself and wait forever. The check now recognizes the whole update chain and watches only the app itself, so updates complete even when antivirus steps in.
• ImprovedIf you are updating from an older build (0.1.11 to 0.1.19) on such a machine, that one hop still runs the old updater and can hang. The escape is simple and one-time: uninstall SideSense from Windows Settings, then install the current version from the download page. Your sessions, settings, and downloaded models stay on your PC through the reinstall.

Canceling or a failed charge no longer locks you out of time you already paid for.

• FixedCanceling a paid plan now keeps the plan active until the end of the period you paid for, exactly as the cancel screen promises. Previously the app could lock the moment the cancellation was recorded.
• FixedA failed renewal charge no longer locks the app on the spot. You now get a grace window of about ten days with a clear "Payment issue" notice while you fix your payment method; the app only locks if the charge stays unresolved past the window.

Reasoning models stop losing their answers to their own thinking, recaps fail honestly instead of saving blanks, and the installer no longer freezes at Finish.

• FixedCut-off or empty answers from reasoning models (like gpt-oss on Groq or OpenRouter). The response budget used to cover the model's hidden thinking too, so a long think left no room for the answer you actually see. The budget now sizes only the visible answer, and thinking gets its own headroom.
• FixedRecaps that came back empty used to be saved as a blank brief with no explanation. A recap that fails or produces nothing now says so on the session page, where you can retry it, and recaps on reasoning models get a large enough budget to finish in the first place.
• FixedThe installer could freeze on its Finish page while antivirus software scanned the app during first launch. Launch is now handed off to Windows, so the wizard closes immediately and the app opens once the scan completes.
• ImprovedPlan pages now clearly mark hosted AI requests, hosted transcription minutes, and web searches as coming soon. Those quotas belong to the paid tiers but the hosted lane has not launched yet; everything running on your PC stays unlimited and uncounted.

Hindi transcription arrives, the live session gets real mid-call controls, and sessions name themselves.

• NewHindi transcription. Pick Hindi under Settings, Transcription (one ~523 MB download, then everything runs offline on your CPU) and sessions transcribe Hindi speech with a model purpose-trained on Indian voices. English stays the default; more languages will unlock the same way.
• NewSwitch the transcription language mid-session. A language selector in the live session's top bar swaps the engine on the fly, so a call that drifts between English and Hindi keeps a usable transcript.
• NewChoose whose questions get answered, mid-session. A new top-bar control switches between both sides, you (mic), or the other side without touching your saved defaults; every session still starts from the default you set in Settings.
• NewSessions name themselves. A session that ends still called “Session 3” gets a real name drafted from what was actually discussed, seconds after it ends. A name you typed yourself is never overwritten, and if naming fails the session page says so and offers a retry.
• Fixed“Name generation failed: The read operation timed out” on slower machines. Naming used the short time budget meant for live answers; it now gets the same generous budget as recaps, so it completes even when the local model has to load from cold.
• ImprovedThe live session's top bar no longer shows the audio-source picker. The source is locked once a meeting starts, so the control only took space; picking the source stays in session setup, where it actually works.

OpenRouter works now, connection tests mirror real sessions, and a custom question detector stays chosen.

• FixedOpenRouter connections work. A timing fault gave every OpenRouter request a fraction of a second before giving up and silently retrying, so tests and answers through OpenRouter never completed. Requests now get the time they need and fail honestly when something is really wrong.
• FixedTesting a provider connection now uses the same settings a real session would, so a model that requires reasoning no longer fails the test while working fine in sessions.
• FixedChoosing the default local model as your custom question detector now sticks. Previously an old one-time settings migration could re-run on the next start and quietly switch the detector back to following your answer engine.

Back/Forward remember your path through Settings, connection tests answer quickly, and saved-confirmations point at the right thing.

• FixedThe Back and Forward buttons now step through Settings pages in the order you visited them, instead of jumping straight out of Settings.
• FixedTesting an AI provider connection no longer spins for minutes when the provider is slow or unreachable. The test makes one bounded attempt and reports the real result within its time limit.
• FixedThe green 'Saved' confirmation in Settings now appears next to the setting you actually changed, not the first one in the list.

One transcription model that just works everywhere, a simpler Transcription page, and working Back/Forward buttons.

• ImprovedTranscription is now one built-in model that runs on every PC's processor, faster than live speech. No presets to compare, no GPU settings to think about - the Transcription page is down to the model, its download, and your language.
• ImprovedSetup is shorter: the Downloads step no longer asks you to pick a transcription model, and the whole first-run download is about 1 GB.
• FixedThe Back and Forward buttons in the title bar now work. They were stuck greyed out even when there were pages to go back to.
• FixedIf you had previously selected the GPU transcription preset, the app moves you to the built-in model automatically and tells you so - nothing to reconfigure.

Long recaps and reports finish reliably, live scoring stops crying wolf, and your saved profile is protected during setup.

• FixedMeeting recaps, interview reports, and follow-up emails on slower engines no longer cut off with an error. The safety timeout added in 0.1.12 was sized for short live answers and could interrupt longer generations; long-form work now gets the time it needs while staying bounded.
• FixedLive scoring no longer shows a false 'failing' warning at the quiet start of a session. Waiting for enough transcript is expected, and the panel now treats it that way.
• FixedFinishing setup can no longer overwrite a saved profile with blank fields when your account is slow to respond; the wizard now waits for your existing profile and keeps whatever you type on top of it. If the profile cannot be loaded at all, the step says so.
• ImprovedHonest states across the board: the Coach voice now explains every fall-back to the browser voice (including when the app itself is unreachable), storage measurements show a real error instead of measuring forever, and a failed size check no longer quietly retries in a loop.
• ImprovedThe built-in voice engine loads strictly from files already on this PC and recovers on its own the moment a voice download completes, instead of waiting out a retry delay.

Fresh installs work out of the box, a faster built-in transcription engine, and honest error messages everywhere.

• FixedFresh installs are healthy out of the box. A packaging fault in earlier builds could leave a new install unable to download its own models or voices; the download engine is now verified end to end before every release, and any machine affected heals automatically on this update, no reinstall needed.
• NewThe Balanced and Efficient transcription presets are now built in, so the Efficient preset runs the moment you install, with only its model weights to download. On a machine with a supported NVIDIA GPU the Balanced preset is available too.
• ImprovedTranscription is English-only for now, with more languages planned; the language setting reflects this honestly. Existing setups are migrated automatically with a one-time note.
• ImprovedQuestion detection now follows your answer engine by default, so a cloud-key setup no longer needs a local model just to detect questions. You can still pick a custom detector engine, and testing a connection now always saves only after it succeeds.
• ImprovedThe installer is smaller overall: the old GPU add-on, which only accelerated the previous transcription engine, is retired, and machines that had it can reclaim the space from Settings.
• FixedEvery model, voice, and connection failure is now visible. Downloads that cannot start, a broken voice falling back to the browser, a device check that could not run, and background size measurements all say what happened instead of failing silently.
• FixedTyped mock-interview practice no longer requires the voice files when the interviewer voice is turned off, and voice add-ons that need an update are surfaced clearly rather than loaded stale.

The update system rebuilt: no more freezes, minutes faster, and nothing fails silently.

• FixedInstalling an update while a session was live could freeze the whole PC. The updater no longer force-stops anything: it waits for the local engine to finish shutting down on its own, and if the engine cannot stop in time the update is called off with a clear message and a one-click retry.
• ImprovedUpdates are minutes faster. The app's bundled runtime now stays in place across updates instead of being reinstalled every time, so a typical update finishes in well under a minute.
• ImprovedUpdating now shows the real installer window with live progress instead of working invisibly in the background.
• NewIf an update fails or is interrupted, the app says so on the next launch and offers to try again. A new Repair installation option on the recovery screen re-runs the last installer, even offline.
• ImprovedThe installer checks free disk space before touching anything, and uninstalling now also removes the app's update download cache.
• ImprovedUpdate downloads announce themselves while they run, and optional add-on packs are re-checked after every update so a stale copy is never loaded silently.

A new transcription engine, sharper question detection, and a big polish pass.

• NewNew transcription engine with two curated presets, Balanced and Efficient, replacing the old model-size list. The Efficient preset runs fast on any machine, no GPU needed.
• ImprovedQuestion detection got smarter: questions split across pauses are stitched back together, incomplete fragments wait until the sentence finishes, and detection keeps listening while an answer is being written.
• NewLive microphone meters everywhere a mic appears, driven by the real input level.
• NewRefreshed window chrome with Back and Forward controls, plus a Display size setting for large screens.
• NewBranded waiting animations with rotating AI facts wherever the app works in the background, and a new startup scene.
• NewSpoken-language setting for transcription.
• ImprovedSimpler settings: friendlier model names, one clear question-detector switch, and a clearer voice-files page with accurate download sizes.
• FixedDark theme no longer resets to light when navigating between pages.
• FixedStarting your very first session from the Home screen now shows the capture notice instead of doing nothing.
• FixedThe voice model card no longer asks you to re-download files you already have.
• ImprovedThe upgrade wizard now says upfront that the previous version is removed first, with your sessions, settings, and models kept.

Critical fix for updates freezing the PC.

• Fixed"Restart and install" could freeze the whole PC while the app shut down. The app now closes its local engine gracefully before an update or uninstall, so updates apply cleanly.

Offline installer, smoother setup, and a transcription fallback fix.

• NewFull offline installer available as a direct download, for machines that cannot reach the update feed or prefer a single complete setup file.
• NewSetup's Downloads step lets you choose the transcription model before anything is fetched.
• FixedWhen a GPU could load the transcription model but not run it, sessions stalled on "Waiting for audio". Transcription now falls back to the CPU automatically and says so.

A lighter installer, faster updates, and a smoother first run.

• ImprovedMuch smaller installer. GPU acceleration and the local voice engine are now optional downloads you add from Settings when you want them, rather than shipping inside the installer.
• ImprovedFaster updates. The app now downloads only the parts that changed instead of the whole app each time.
• FixedNo more large background download with no visible progress right after setup finishes.

Onboarding, live view, and stealth refinements.

• NewGroq added as a bring-your-own-key AI provider during onboarding.
• NewLive view now lists the questions detected in every session type.
• NewAnimated startup splash while the app warms up.
• ImprovedSimpler stealth: a single overlay control, with adjustable window transparency.
• ImprovedBranded installer wizard.
• FixedMock-interview question generation no longer fails when a local Ollama model returns malformed JSON.

First private early-access build for testers.

• NewCoach practice mode with scored mock interviews and a growth dashboard.
• NewAssist live mode: quiet real-time answer drafts from your own material.
• NewFully local lane with Ollama for on-device AI, plus bring-your-own-key providers.
• NewLocal transcription with automatic CPU fallback when no compatible GPU is present.
• NewWindows capture of microphone and system audio (loopback).