TypeWhisper 1.8.3 Release Candidate 1

The first release candidate for TypeWhisper 1.8.3. It contains everything merged since 1.8.2.

Highlights

• Live dictation with more cloud engines: AssemblyAI, Mistral AI, Reson8 and ElevenLabs stream dictations while you speak. The live preview, live text in the text field and the Live Transcript window show the provider's own transcript, and the final text is ready a fraction of a second after you stop.
• Every Recorder transcript is saved to History, not only recordings with speaker detection, so History search, sync and exports include them.
• Push-to-talk keeps recording up to 350 ms after you release the key while you are still speaking, so the last word is no longer cut off.
• Dictations no longer ask cloud engines for speaker labels, which providers bill on top of the transcription and dictated text never shows.
• The Mac App Store edition can redeem App Store offer codes.

Dictation

• After a push-to-talk release, a toggle press or another hotkey stop, the recording continues for up to 350 ms and ends as soon as the last 150 ms are quiet. If you paused before releasing, it ends at once. Very short taps and stops through the API or after silence behave as before.
• Stop recording after silence: an engine that recognizes when you finished speaking, such as Reson8 with its turn detection, can end a toggle dictation after half the silence duration, at least one second. A pause in the middle of a sentence still waits the full duration.
• Keyboard input no longer lags for seconds when key events queue up while TypeWhisper is busy, and a toggle Fn hotkey no longer loses its start or stop after such a stall.
• After undocking a Mac from clamshell mode and sleep, recordings recover from the missing audio device instead of failing until TypeWhisper is relaunched.
• On a MacBook during a call, other apps' audio, the call included, is no longer turned down while TypeWhisper only prepares the built-in microphone between dictations.
• TypeWhisper can read the focused text field in Google Chrome during a dictation, for example for correction learning. Chrome ignores the switch that other Chromium browsers and Electron apps use, so TypeWhisper now falls back to the one Chrome supports.
• With Show live transcript in the active text field, Telegram drafts no longer end up with duplicated text.
• The interface language follows a language set for TypeWhisper in System Settings > Language & Region, so labels and durations no longer mix two languages.

Recorder and meetings

• Every transcribed recording gets a History record with its transcript, name, time, duration and engine; the audio stays in the Recorder's folder. Transcribing a recording again updates its record. Recordings transcribed before this version are imported once. The records need no Premium and sync like other History records. With Save history off, no History record is saved.
• With calls played on speakers, the Recorder no longer credits the other participants to "Me". Your own speech now has to be clearly louder than the echo of the call on the microphone, which TypeWhisper measures from the recording.
• Recordings transcribed with an engine that reports word timestamps get <recording>.timestamps.json next to them. Deleting a recording moves it to the Trash with the other files.
• Watch folders: the Markdown transcript starts with frontmatter that names the audio file, its word timestamps, the date, length, language, engine and speakers, and has a paragraph per speaker turn that starts with its time and speaker. With an engine that reports word timestamps, <file>.timestamps.json holds them in the format of the HTTP API's verbose_json.
• Script Runner 1.3.0: scripts that run for Recorder transcripts of recordings started from the calendar get TYPEWHISPER_MEETING_TITLE, TYPEWHISPER_MEETING_START, TYPEWHISPER_MEETING_END and TYPEWHISPER_MEETING_ATTENDEES.
• Obsidian 1.3.0: Recorder notes name the recording and its word timestamps in their frontmatter (audio:, timestamps:) and show its length.

Plugins and SDK

• AssemblyAI 1.2.0: dictations with Universal-3.5 Pro stream to AssemblyAI while you speak. Recordings, files and dictations without a live session use AssemblyAI's Sync API, which returns the text in one request instead of uploading, queueing and polling. A new Transcription setting shows Live, Sync and Async with AssemblyAI's list prices.
• Mistral AI 1.3.1: dictations with a live preview stream to Voxtral Mini Transcribe Realtime 3, and the transcription model writes the final text. A Transcription setting shows Live and After recording with Mistral's list prices. Transcription uses Voxtral Mini Transcribe 3 by default and Voxtral Small Transcribe 3 for higher quality; the language models are Mistral Large 4, Mistral Small 4 and Ministral 3. Dictionary terms reach Mistral as context, speaker labels mark who is speaking in recordings, the language picker offers 58 languages, and the model lists come from your Mistral account.
• Reson8 1.2.1: the model picker offers Rumble-2-Flash next to Resonant-1 and your custom models, and dictations with a streaming model stream to Reson8. Recordings and files use the prerecorded endpoint, which costs half the credits. New settings cover filler words, dictionary terms as phrases or a custom model in your Reson8 account, pattern sets for structured tokens, and speaker diarization for recordings. End dictation when you finish speaking uses Reson8's turn detection.
• ElevenLabs 1.2.0: in Auto mode, Scribe v2 dictations stream to ElevenLabs while you speak, including dictations longer than 30 seconds. With active dictionary terms the final text still comes from the batch API.
• Microsoft AI 1.1.0: an optional backup endpoint with its own API key takes over when Azure Speech refuses a request for lack of capacity in the first region. Audio goes to MAI-Transcribe as FLAC, and the regions centralindia and westus2 no longer show the region warning.
• OpenRouter 1.2.0: MAI-Transcribe models receive FLAC instead of WAV, about half the size, so long dictations no longer slow down above about 110 seconds of audio.
• Soniox 1.2.15: dictation and file transcription use Soniox STT v6, which Soniox released on November 3, 2026; it retires v5 on December 15, 2026. A pinned v5 model moves to Automatic (Latest).
• Speaker diarization in Reson8, AssemblyAI, Microsoft AI, Mistral AI and Meta applies to recordings, file transcriptions, watch folders and the HTTP API only. With AssemblyAI, dictations then use the Sync API.
• Granite Speech 1.2.2 frees TurboCTC's memory after each transcription and on unload. Before, memory grew by about 12 MB per recorded second until TypeWhisper restarted.
• Local LLM (MLX) 1.0.8, Qwen3 ASR 1.1.16, Granite Speech 1.2.2, Voxtral 1.0.21 and Canary ASR 1.1.2 use swift-transformers 1.4.0, which updates yyjson to 0.13.0 and fixes a stack buffer overflow with crafted JSON files larger than 4 GiB (GHSA-f4vc-345x-4mvm). Local LLM keeps the spaces Gemma models write before punctuation.
• Plugin SDK: PluginSpeakerLabels.areRequested tells an engine whether the host uses speaker labels for the current transcription. Engines opt in to ending dictations at turn ends through the Objective-C method setLiveTranscriptionTurnEndHandler:, so their plugins stay loadable by older versions.

Mac App Store edition

• The Premium page shows Redeem Code next to Restore Purchases on macOS 15 and later. It opens Apple's offer code sheet inside the app, and Premium unlocks right after the redemption. If the App Store transaction cannot be verified, the page says so.

Fixes

• Restarting from Update All and Restart on the Plugins page or from the language setting on the General page reopens Settings on that page.
• The About panel of Daily and Release Candidate builds names the channel, for example "Version 1.8.3 Daily" followed by the build number.
• When Apple Intelligence on macOS 27 rejects a text as too long for its context, the local API answers 413 context_length_exceeded instead of 503, so a paired iPhone retries with a shorter excerpt.

Compatibility and updates

Requires macOS 14 or later; supports Apple silicon and Intel Macs. The iCloud bridge is included in every release.

Reson8 1.2.1, AssemblyAI 1.2.0, Microsoft AI 1.1.0, Mistral AI 1.3.1 and Meta 1.1.0 require TypeWhisper 1.8.3; earlier versions keep the plugin versions they have.

RC1 is available through GitHub and the Release Candidate update channel. The stable release and Homebrew remain on 1.8.2. Please report regressions with the release tag, macOS version, and the affected engine or plugin.