High-accuracy speaker labeling can now be turned off from Settings or the Transcription menu. Silo Scribe goes back to its offline speaker estimate and deletes the downloaded speaker model from your Mac; labels already in your transcripts stay as they are. If the speaker model you chose can't be loaded, Record and Settings now say so and offer to download it again or turn it off. Declaring, revising and revoking clinician authorship in Linked Evidence now work; before, each attempt failed with an error. Silo Scribe also has a new app icon.

New: choose your transcription engine. Use Apple's on-device Speech framework (no download), or an optional local MLX speech model (Qwen3-ASR, about 1 GB, downloaded once after explicit consent from a pinned, checksum-verified source) for both recorded sessions and Option-Command-D dictation. A session whose transcription cannot run is now kept and retryable instead of being discarded. The model is repairable and removable in Settings, engines never switch silently, and no audio ever leaves your Mac.

New: choose your transcription engine. Use Apple's on-device Speech framework (no download), or an optional local MLX speech model (Qwen3-ASR, about 1 GB, downloaded once after explicit consent from a pinned, checksum-verified source) for both recorded sessions and Option-Command-D dictation. A session whose transcription cannot run is now kept and retryable instead of being discarded. The model is repairable and removable in Settings, engines never switch silently, and no audio ever leaves your Mac.

Introduces on-device Clinical Billing Assistant with suggested ICD-10 diagnostic categories, CPT evaluation & management code references, and one-click formatted EHR pasteboard export desks. All processing runs 100% locally on Apple Silicon with full zero-knowledge privacy.

This update fixes four ways a session could go wrong without Silo Scribe telling you. The input list now shows only devices that can actually record. Speakers, multi-output devices, and the playback side of a USB interface no longer appear, so a session can no longer be captured as silence -- and if your chosen microphone cannot be selected, recording refuses to start instead of quietly recording from a different one. Stopping a recording now waits for the full on-device transcription however long it takes. Long sessions previously ended with an empty transcript and no explanation. A recording that ends because a microphone was unplugged now shows its transcript right away, and anything that does go wrong is reported instead of failing silently. Transcription now falls back to English when your Mac's region has no on-device speech model installed, instead of refusing to transcribe at all. Notes are safer at rest: the encryption key is never replaced when the Keychain simply cannot be read, so existing notes can no longer be left unreadable. Everything still runs entirely on your Mac.

This update fixes the microphone and speech permission flow on a fresh install: Silo Scribe now waits for your answer to the system permission prompt instead of reading a stale status, so the record button enables as soon as you grant access. Note generation can no longer hang indefinitely -- the local helper connection now has time limits, so a stalled model reports an error instead of sitting on "generating..." forever. Quitting the app no longer beach-balls while the helper shuts down. Everything still runs entirely on your Mac.

Fixes a defect that prevented recording from starting on a new install: the microphone permission prompt could not be reached, so the record button stayed disabled. Silo Scribe now asks for microphone access up front and offers a direct route to System Settings if access was previously denied. Model output that cannot be parsed into a structured note is now clearly marked as unstructured instead of being presented as a finished note, and a failed patient-history search now reports the failure rather than showing it as no matches found. Everything still runs entirely on your Mac.

Create private dot phrases and custom clinical note structures from wording you author, and insert them into the exact section you choose. Dictate sessions with on-device transcription and review every line before drafting. Note generation now runs the on-device Gemma 4 model reliably, with clearer manual-versus-generated labels and encrypted local storage for reusable note tools.

Full light and dark mode following your system appearance. Transcript review workshop improvements: click any line to correct it, reassign or rename speakers, merge fragmented lines, and see per-speaker talk time — with low-confidence lines flagged for proofing. Custom vocabulary: add your clinical terms in Settings, or promote a corrected term straight from the transcript. Notes drafted without the local AI model are now clearly labeled, with a link to download it. One-click Copy note for pasting into your EHR. A speaker-diarization badge on the Record screen shows when speaker separation is active.

Speaker labeling (high-accuracy diarization) is now available in Settings. Note generation shows live status and fails visibly instead of hanging. Recording now recovers cleanly if your audio device changes mid-session. First launch guides you straight to adding a patient.

New: dictation hotkey (⌥⌘D) to capture notes hands-free with on-device speech recognition — nothing leaves your Mac. Reliability and performance improvements throughout.