What's new in AnyBrief

Smaller recordings and more flexible exports

Export rules can now save the summary, the transcript, or both documents.
Filename templates support a new {type} variable that expands to summary or transcript.
The meetings list now includes an action to reveal exported files in Finder.

Raw meeting audio is now compressed to MP3 even when summarization is not configured or does not complete, significantly reducing storage use.
The local meeting recording remains the source of truth, so transcription and summarization can still be repeated as before.

FluidAudio offline diarization now uses the full OfflineDiarizerManager pipeline.
Speaker count supports either an exact value or an upper bound; calendar automation uses the maximum expected participant count.
FluidAudio and whisper.cpp now have separate recognition dictionaries that follow the selected STT provider.

Whisper now uses Silero VAD to remove silence before recognition, suppressing invented phrases and subtitle-credit hallucinations.
Original word and speaker-turn timestamps are restored correctly after VAD processing.
Diarization models now have a dedicated preparation command instead of processing a silent WAV.
The recognition dictionary now sits with the selected STT options, and the provider selector is larger and clearer.

The shared recognition dictionary is now passed to FluidAudio and whisper.cpp, with preferred spellings and known variants.
Speaker separation can be disabled for faster recognition without diarization.
Meeting actions now include Repeat Transcription and Repeat All.

FluidAudio recognizes the full recording in one pass and aligns timestamped words with speaker turns, preserving context across long meetings.
A fixed speaker count now constrains diarization output, and the default speaker sensitivity is now 0.65.
Repeat transcription now converts archived MP3 tracks to temporary 16 kHz mono WAV before FluidAudio or whisper.cpp runs, preventing MP3 decoder errors and hallucinations.
Recognition logs now report the actual provider, model, and language instead of placeholder values.
Processing status now shows transcript cleanup, the active LLM connection, and fallback transitions.
Before launching Codex or Claude, the app quickly checks its API and immediately falls back to the next LLM when the service is blocked or unavailable.
CLI providers now use only the connection's overall timeout, so long generations without intermediate output are no longer stopped after 120 seconds.
Microphone, speaker separation, and recognition dictionary settings now use compact collapsible cards.
The provider selector and whisper.cpp options now follow one consistent grid, with larger controls and tighter grouping.
Russian and English localization has been reviewed and completed.

A second transcription provider: local whisper.cpp recognition combined with FluidAudio speaker diarization.
Choose between tiny, base, small, medium and large-v3-turbo-q5_0 Whisper models, plus a specific recognition language or automatic detection.
Metal acceleration and provider-specific diagnostics for Core ML, Apple Neural Engine, Metal, command-line tools and model availability.
Summary frontmatter now records the transcription provider, model, language and acceleration mode.
Choose a specific microphone or follow the current macOS system input automatically.

Model and provider technology cards in Transcription settings are now collapsible and use much less space.
System-audio and microphone meters now share a consistent grid.
Russian and English localization has been reviewed and completed for the new settings.
If FluidAudio finds no speaker turns, the recognized text is now preserved as one segment instead of failing with an Invalid combined.txt header error.
Stalled CLI summarizers that produce no output are stopped after 120 seconds together with their child processes, then the next LLM in the fallback chain is used automatically.
The release includes universal binaries for Apple Silicon and Intel Macs.