A plain-language record of what's changed in IWV Chat AI with each release — what we fixed, what we improved, and what's new. The same information is available inside the app under Help → Release Notes & Errata.
Fixed: In Conversational Mode, if you press Done Speaking (or your pause is detected) and no words could be understood, the app now says so out loud instead of silently listening again — previously a microphone problem could look like the app simply ignoring you. Quiet waiting periods are unaffected.
Fixed: On Android, the status display now updates to “Working…” immediately when your speech is being processed, instead of appearing frozen on “Listening…” until the answer was ready.
Added: Conversational Mode on Android now detects when you stop speaking, just like on desktop — your words are processed after your natural pause instead of waiting for the 12-second listening window or the Done Speaking button (both still work as backups). Detection adapts to your device’s microphone and background noise, and calibrates itself to your voice over the first few turns.
Improved: Voice transcription on Android now runs only on the phone’s performance cores instead of all cores. On devices that mix fast and efficiency cores, this makes speech-to-text noticeably faster — the same optimization the AI engine already used.
Fixed: On Android, the AI model could be loaded into memory twice when two parts of the app warmed the engine at nearly the same moment (for example, app startup and entering Conversational Mode). The duplicate copy could not be freed, roughly doubling memory use — on lower-memory devices this could make the microphone go silent during voice conversations or cause the system to close the app. Model loading is now strictly one-at-a-time, and a second request simply reuses the load already in progress.
Fixed: The app icon could show black corners around the circular logo instead of transparency — on Windows (a malformed icon file that some renderers mis-parsed) and on Android (launcher icons exported with opaque black corners). All icon assets regenerated with verified transparency at every size.
Improved: Reporting an AI response now always includes the correct originating question and notes any image that was part of that exchange — including follow-up answers about an image attached earlier in the conversation (marked as such). Only attachment names and types are included, never the files themselves.
Improved: Very large images (30 MB and up) are now processed with a memory-safe method adopted from NotesXML: the image is decoded and scaled directly, without building huge intermediate copies — large panoramas and scans attach faster and more reliably.
Fixed: Images added by drag-and-drop or paste were being scaled to a smaller size than the vision model expects; they now use the correct per-model resolution (matching images added via the file picker).
Fixed: Attaching a large image could fail silently — nothing appeared and no reason was given. Failures now post a notice directly in the chat, and the image size limit has been raised from 10 MB to 100 MB (images are scaled to the AI model’s input size regardless, so ordinary large photos, panoramas, and scans now simply work).
Fixed: Spoken replies could clip the first word or two when speech started after a quiet period; a brief silent lead-in now wakes the audio output first, independent of your pause setting.
Improved: Much shorter gaps between spoken sentences: the voice engine now stays loaded for the whole session instead of restarting for every sentence, and the between-sentence pause setting now controls exactly that pause and nothing else.
Fixed: The AI engine was being launched with a far larger context window than intended, allocating a very large memory cache that slowed response generation — most visibly on integrated-GPU systems (roughly 4x slower than the hardware can deliver). The intended context size is now enforced.
Added: The AI engine now runs on Intel Arc/Iris integrated GPUs on Linux (previously CPU-only on those systems), with roughly 3x faster responses on tested hardware. The underlying inference engine was updated accordingly.
Fixed: Images sent to the AI were being downscaled below the vision model’s native input size, reducing OCR and image-understanding quality. Vision inputs now use the correct per-model resolution.
Fixed: On systems where the AI was set to CPU-only, background GPU activity could still occur and slow the machine; CPU-only now fully disables the GPU path.
Fixed: Live voice-activity sampling could read stale audio on systems where the recorder writes in large bursts, disabling automatic end-of-speech detection. Recording is now written at fine granularity and only fresh audio is analyzed.
Changed: Voice Settings now apply immediately as you change them. The Save button has been removed — previously it sat below the fold and changes were silently discarded if it was missed.
Added: The in-app Release Notes & Errata page.
Added: Modulation-based voice detection for microphones whose processing (automatic gain control / noise suppression) flattens loudness so that level-based detection cannot work. The right detector is selected automatically per machine.
Fixed: Voice detection could lock into a “speech” state under steady background noise (for example a cooling fan), preventing automatic end-of-speech; the adaptive noise floor was redesigned (1.7.0).
Fixed: The inactivity timeout could end Conversational Mode while an utterance was still being processed, discarding it (1.7.1), and could end an active conversation after 2.5 minutes on systems without working end-of-speech detection (1.7.3).
Added: Per-machine speech-level auto-calibration for voice detection (1.7.2).
Fixed: Silence or background noise could occasionally be transcribed and sent to the AI as a phantom turn (1.6.0); an over-aggressive background-audio guard could exit the mode and discard a valid spoken question (1.6.2); idle sessions previously never timed out (1.6.2).
Added: Conversational Mode safeguards: entry requires a deliberate mouse/touch press; opening the camera or typing exits the mode; continuous background audio (TV, music) no longer keeps it running indefinitely (1.6.1). The AI engine is warmed when entering the mode, making the first spoken answer much faster (1.6.0).
Fixed: Text-to-speech could speak formatting symbols (“asterisk asterisk”) and wasted synthesis on punctuation-only fragments (1.5.1).
Added: Persistent speech-recognition engine: the transcription model loads once per session instead of once per utterance, removing a fixed delay from every voice turn (1.5.0).
Fixed: End-of-speech detection previously depended on transcription timing, so the “pause after you stop speaking” setting did not measure your actual pause, and long utterances could be cut off mid-speech.
Added: Real-time voice-activity detection: the pause setting now does exactly what it says.
Added: Voice model selection with a downloadable catalog of additional voices, multi-speaker voice support, and a Test button (1.3.0–1.3.1).
Fixed: The spoken “Processing” cue delayed the AI request instead of playing alongside it (1.3.5); light-theme title bar contrast (1.3.4); settings styling issues (1.3.2–1.3.3).
Added: Conversational Mode tuning controls: adjustable pause after you stop speaking, adjustable pause between spoken sentences, adjustable live-transcription chunk period.
Added: Android 15 compatibility (16 KB memory-page devices) and updated target SDK (1.1.0).
Changed: Application identifier finalized for store distribution (1.1.1); launcher icon corrections (1.1.2); AI model catalog update (1.1.4); desktop GPU backend housekeeping (1.1.5).
This page currently covers releases 1.1.0 through 1.17.1; notes for the most recent releases are being finalized. Have a question about a specific fix? Contact support.