Updates

Talkaside 1.0.5: faster model selection and better answers

23 October 2026 · Development update · Work since 16 October

The 1.0.5 installer is now available from the homepage.

• Auto picks a model that passed its test Test the available models and Talkaside records their response times. Auto uses the fastest model that passed, and keeps the results after a restart. Models that fail stay visible in the list, greyed out with a reason such as a timeout, rate limit or missing quota.
• Auto can start a backup before the first provider finishes When AI1 uses Auto, Talkaside starts a backup request if the first provider has not responded after two seconds. It uses the first answer that passes the checks, within a 12-second response window.
• Answers stay closer to the facts in your CV Talkaside now carries details such as education, salary expectations, location, relocation plans and English level into relevant questions. It checks proposed answers against those details, handles percentages from the CV, and avoids filling gaps with claims the document does not support.
• More spoken questions get a direct answer Phrases such as “Can you…” and “Tell me about…” are now recognised as questions, including when they appear inside a longer turn. The answer is matched to what was asked instead of defaulting to “I haven’t had that experience.”
• Provider problems are easier to recover from Talkaside no longer opens the replacement prompt when it rejects an answer for a CV conflict. If a provider hangs through two consecutive requests without producing words for 15 seconds, it offers a replacement. A provider cooldown triggers one retry rather than an endless loop.
• OpenCode and provider tools work more reliably Free OpenCode models can answer again, and the OpenRouter model list now loads on slower connections. Settings include Update controls for Claude, Grok and OpenCode.
• Your network settings can be reused Save custom DNS-over-HTTPS servers and proxy connections in Talkaside settings. The app keeps those choices for later sessions without changing Windows-wide network settings.
• Talkaside can install the next release from the app The Update button downloads the next release with progress, checks its signed release data and SHA-256, then runs the installer over the existing version. Install 1.0.5 from the homepage; the in-app updater handles later releases.
• Release builds expose less build detail Hardened the packaged Electron build and removed repository documentation and local build paths from the release package.

Talkaside 1.0.4: redesigned interface and local models

15 October 2026 · Changes from 8 to 15 October

• Redesigned app interface Brought the main window, settings, history, teleprompter, floating panel and activation screen into a consistent design. All windows now use Segoe UI, including buttons and input fields. Improved label contrast, spacing, selection highlights and control states.
• Reorganised settings and dropdown menus Reorganised speech recognition, translation and voice output settings. Removed repeated explanations and unnecessary buttons. Local components now have separate Install and Test controls. Fixed menu positioning in small windows, scrolling in long lists and keyboard navigation.
• Updated installer and uninstaller Updated the layout of the licence agreement and confirmation screens, and added uninstall animations. Downloaded local models can be removed through a separate option. Settings, history, recordings and My Voice samples are kept.
• Expanded language support Speech recognition supports more languages through cloud providers and local models. The available languages depend on the selected engine. Language names appear in their native form; search accepts Russian and English names and language codes. Regional variants are preserved.
• Improved AI answers during calls Improved detection of a completed question while the other person continues speaking, and stopped already processed questions from being sent again after final transcription. Added checks for clear contradictions with the age or place of residence in your CV. Fixed the second-answer waiting indicator: it disappears when text starts arriving. The AI answer language can be set separately from translation.
• Fixed spoken questions after screenshots Spoken clarifications close to the moment of capture are considered alongside the image. Later speech no longer extends that window indefinitely. Fixed a case where new speech kept restarting or cancelling a screenshot answer before it was ready.
• Improved voice output controls Voice selection and on/off controls now stay in sync between settings and the main toolbar. Fixed voice previews, including when call voice output is disabled. Playback failures now show an error instead of a false success message.
• Install My Voice components from the app Download the required dependencies with the Install button in voice settings. There is no need to install Python or run commands manually. Large components and models are not bundled with the main installer. Local voice cloning requires a compatible NVIDIA GPU; your sample and voice profile stay on your computer.
• Added local models Whisper and Parakeet are available for speech recognition, Hy-MT2 for translation, and RapidOCR for extracting text from screenshots. Install and test the components you need from settings. Recognition, translation and OCR can run locally; whether the resulting text is sent to AI depends on your selected provider.
• Fixed recording controls and improved history Automatic call recording requires explicit opt-in. Call REC is shown while audio is being saved; changing the setting during a call applies to the next call. Simplified how time segments are displayed and played in history. Fixed handling of a recording file becoming unavailable while it is being saved.
• Included a local question-and-answer bank The app includes 126 entries in Russian and English covering AppSec and interview preparation. Local search works without the original database folder. Your own documents and personal question banks are not included in this shared bank.
• Checked security and improved app protection Also improved Windows startup and error diagnostics.

Local processing, call review and setup

8 October 2026 · Development update · Changes since 7 September

This entry describes development as of 8 October. See the 1.0.4 release above for the features included in the published demo.

• Interview answers and preparation Updated question detection, follow-up handling and matching to saved answers. Import questions and answers from documents, or load a job description from a link. Added checks for personal facts that conflict with the supplied CV; AI answers still need checking.
• Screenshots and teleprompter Choose a monitor, window or region for a capture. Local OCR can extract text; vision models can analyse the image. Extracted text may still be sent to online AI. Captures can provide context for later spoken questions. The teleprompter follows your voice, lets you jump to a word and accepts AI answers.
• Speech recognition and local models Added Speechmatics, local Whisper and Parakeet, alongside Deepgram and Google Gemini. Whisper offers CPU, CUDA and Vulkan backends; Parakeet offers CPU and Vulkan. Local runtimes and models are optional downloads, with separate Install and Test controls. Recognition can switch to configured backup providers after a failure.
• Translation, voices and language practice Added optional local Hy-MT2 translation. Auto tries DeepL, Google Gemini, Groq and installed Hy-MT2 in that order, skipping unavailable options. Language choices depend on the selected providers. Added separate voices by language, an ElevenLabs model selector and sentence-by-sentence playback. Language practice offers suggested replies for continuing a conversation.
• Recording, imports and call review Choose audio, screen video or both when recording. Import WAV or MP3 files for local transcription when no live session is running. Search transcripts, play recordings, add notes and request a summary or interview review from your selected AI. Transcription revisions are kept separately. Imported originals and exported files are not deleted with a stored call.
• Provider selection and first launch The setup guide checks speech access, AI access and a trial question; adding a CV or job description is optional. AI can connect to an existing local Ollama or LM Studio server. Model lists use provider catalogues and access tests. Added connection diagnostics, optional Shadowsocks, SOCKS5 and HTTP proxies, and settings controls for updating Codex CLI and Claude Code.
• Licensing, settings and crash recovery Prepared the five-day demo and updated file-integrity and licence-expiry checks. Updated settings, waiting indicators, notifications and the Bluetooth headset display. Electron was updated from 43.3.0 to 43.7.7 after a crash was found, with window recovery and diagnostic logging added.