MachinesFluent v1.2.5 adds a Transcript Library, Google audio and video transcription, editable AI summaries, and an Ask Assistant action for saved transcripts. Voice Notes now displays formatted pages, Quick Correction lets you choose a dictionary for new words, and the update refreshes speech and AI choices with clearer errors and safer saving.
Transcribe files and keep the results in the app
Open Settings > File Transcription, add your audio or video files, and choose how to transcribe them. You can use your active local Whisper, Parakeet, or Cohere model, or select Google Gemini 3.5 Transcribe independently of the model you use for live dictation. The Google option needs Pro access and your existing Google API key; it does not require a local speech model to be loaded.
Both routes offer four output choices: plain text, timestamps, speaker labels, or timestamps with speakers. Google handles file uploads, including extracting audio from video and splitting long recordings into smaller uploads. Local transcription keeps the speech processing on your computer. Google transcription sends the audio to Google, so choose the route that fits the recording.
Completed files now appear in Transcript Library. Open a transcript to read it, search with highlighted matches, edit passages, change its title, copy its text, or delete the saved result. Renaming or deleting a transcript leaves the original audio or video alone. Saves also check for newer edits before replacing existing text.
The library replaces the extra automatic TXT copy and destination-folder setting for new transcriptions. Your previously created files are preserved, and you can still copy a transcript or open its saved-file folder. If you are starting with private recordings, the local transcription guide explains the offline workflow.
Name speakers and create a summary you can reuse
Local speaker identification now supports up to eight anonymous speaker channels. It requires compatible DirectML hardware and a separate download the first time you use it. These are speaker labels, not automatic recognition of real people. Google provides its own labels for each uploaded piece; a label in one piece does not establish the same person's identity in another.
You can assign names to speakers and keep a reusable name list. Those names appear in the transcript and its summary. Reset to Default restores a speaker's numbered label without deleting the reusable names you have saved.
Open the transcript's AI Summary tab, then use Summary Settings to choose a provider, model, and format:
- By topic groups the summary around the subjects discussed.
- By speaker organizes it around the identified speakers and is available when the transcript has speaker labels.
- Custom uses your own instructions for the result you want.
All three formats have editable instructions, and each can be restored individually. For example, a Custom instruction could ask for decisions, open questions, and action items from a meeting. The summary remembers the format, model, provider, and instructions used, so changing the current settings does not mislabel an older result.
Editing transcript text marks an existing summary as outdated. Regeneration keeps the previous summary if the request fails, and unsaved instruction drafts survive navigation and rejected saves. Summary generation uses the chosen AI provider, which can be local or cloud; local transcription alone does not make a cloud-generated summary offline. Check important conclusions against the transcript before using them.
Ask Assistant about the transcript
Use Ask Assistant in the transcript reader to attach the complete current transcript and leave the message box ready for your question. An existing conversation and unsent draft stay in place. You could ask which decisions are still unresolved, request an email based on the discussion, or ask for the passages that support a particular point.
The attachment is a snapshot of the transcript when you add it. Later edits in the library do not update that attachment; add the revised transcript again if you want Assistant to use it. The selected model's normal attachment limits still apply, and sending to a cloud AI provider sends the attached text to that provider.
Assistant also keeps newly added files and message drafts after a rejected send, gets more vertical room, and renders Note and Warning boxes more clearly. New settings use an 80-second AI request timeout instead of 40 seconds; an existing saved timeout is preserved.
Read Voice Notes and fix vocabulary more easily
Voice Notes now displays saved notes as formatted Markdown pages inside Settings. Search, copy, or delete notes from the library, and switch to the raw-text editor when you want to change the content. The stored note remains a local plain-text file. You can also edit the title directly and press Enter to save it without discarding an unsaved content draft.
Rejected or unconfirmed saves preserve the note draft and its identity rather than creating a replacement note. This is useful when a connection drops before a save is confirmed: the text stays available while you check the refreshed library.
Quick Correction now lets you choose where a new word goes: the Quick Corrections dictionary or an existing dictionary. It preserves that dictionary's enabled state. The compact window can be dragged, follows your appearance settings, and shows a check mark after saving.
Word Replacement Strictness now lives in Vocabulary, with one shared slider above Dictionaries and Snippets. The update also fixes exact Voice Snippet expansion and cases where more than one dictionary correction belongs inside a single word. If you missed the original shortcut, the v1.2.4 release post explains Quick Correction and one-recording voice triggers.
New speech and AI choices
Microsoft MAI-Transcribe 2 is now available for dictation through OpenRouter using the same OpenRouter API key. It transcribes completed speech after pauses and uses the provider's clean transcription style. This dictation choice is separate from the new Google file-transcription option.
Local Whisper, Parakeet, and Cohere also get a Force CPU choice under Experimental ARM Support / Force CPU in General Settings. Restart MachinesFluent to apply it. Despite the experimental label, this selects CPU processing; it does not add native ARM or NPU support.
AI setup now includes a DeepSeek API-key walkthrough and a DeepSeek Flash Non-Reasoning choice. The shipped model menu has been refreshed with GPT-6 Sol, Gemini 3.5 Flash-Lite and 3.8 Flash, Claude Opus 5.5, MiniMax M3, and GLM 5.3. Supported saved choices remain yours, and the speed, intelligence, image, and search badges have been updated.
Assistant offers more web-search routes through OpenRouter and supported models from Alibaba, xAI, Moonshot, Groq, and DeepSeek. Availability depends on the selected model and provider; a search badge does not promise that every answer will run a search.
Mistral and Perplexity AI support and Gladia speech dictation have been removed. If you used one of those providers, choose a supported replacement after updating. Selected models and active Smart Action or Smart Dictation prompts are now protected from removal until you select another choice.
Everyday fixes and interface improvements
- File transcription shows clearer progress and processing stages, with better handling of long files and cancellation.
- Speech-model loading, switching, downloads, and shutdown are safer. Parakeet keeps meaningful interjections such as “oh” and “ooh,” and Speechmatics connection and failure handling has improved.
- Failed computer-sound capture stops with an error instead of unexpectedly recording the microphone. Leftover sentence counts and recording feedback are cleared before the next session.
- AI and cloud speech errors give more specific explanations for unavailable models, account restrictions, provider limits, connection problems, and empty replies.
- Settings has more consistent editors, dialogs, search fields, copy controls, help, save feedback, and scrollbars. Transcript editing keeps a stable frame, and tutorials are easier to navigate.
- The system tray now offers Top and Bottom dock positions. New settings use Matrix Ripple and Dock Glow, while existing appearance choices are preserved.
- Website detection failures are contained so they cannot take down dictation, and malformed hotkeys or numeric settings no longer discard unrelated preferences.
- The new controls and messages are translated across English, French, Spanish, Italian, German, Russian, Japanese, and Korean.
Update to v1.2.5 to keep file transcripts, summaries, follow-up questions, and voice notes together in MachinesFluent.
Download MachinesFluent for Windows.
