Draft transcript
Clean local pass
Summarize the optic nerve finding and prepare a concise handoff.
Temporal pallor noted; correlate with OCT and visual field trend.
Words today
18,420614 training-safe corrections promoted locally.
WPM avg
137Confidence drift -4% after correction lift.
CaptureHold-to-dictate
Voice profileOwner verification
CorrectionsLearned replacements
Prompt drillsFormatting rehearsal
Rewrite reviewFinal wording
Voice inputSpectrographTokensTranscriptOutput
Live signal context
The app keeps the signal, draft, and approved learning boundary in one readable pass.
Section two now shows the practical operating loop: deliberate capture, local transcript, confidence/pace telemetry, and only the corrections you approve becoming reusable local context.
Signal telemetry
Confidence stays visible
137 WPM94% clean route-4% drift
Learning boundary
Approved corrections only
rapid focus leg -> rapid refocus lag
Nothing enters the local profile until accepted.
End-to-end signal path
One signal, one visible route from sound to finished text.
A spoken prompt enters as waveform energy, becomes a spectrogram, resolves into likely tokens, passes through the local signal chain, and lands as useful writing.
CoreScribe signal consoleVoice prompt -> waveform -> spectrogram -> tokens -> output
01 Waveform
"Summarize the optic nerve finding."02 Spectrogram
Sound becomes an image the model can read.
high pitchlow pitchtime
03 Tokens
Likely language starts forming.temporalpallornotedcorrelateOCT
04 Signal chain
Encoder memory and decoder context align.05 Output text
Three prompt styles from the same captured signal.Clinical noteTemporal pallor; correlate with OCT/VF trend.AI promptAsk duration, laterality, triggers, red flags.MessageRunning late; I will call when I park.
Processing path
How CoreScribe turns voice into finished text.
The product path is the point: voice input becomes spectral evidence, the local model proposes tokens, rewrite logic shapes the draft, and the finished output lands in the focused Mac app.
CoreScribe processing path
Voice inputSpectrographTokensTranscriptOutput
Writing
LocalVoice to finished output
How it works
A local transcription path you can actually describe.
CoreScribe combines intentional capture, on-device speech recognition, rewrite controls, correction memory, and focused-app insertion. The point of this page is not mystery; it is to show what the app is actually doing.
01
Intentional Capture Starts the Pass
The hotkey opens a deliberate recording window instead of an always-listening microphone. CoreScribe records only during that hold, then hands the captured audio to the local transcription path.
02
Audio Becomes Local Text
The audio is turned into model-ready features, the local engine proposes tokens, and the transcript is assembled on-device. No cloud service is required for normal dictation.
03
Modes and Corrections Shape the Draft
Voice profiles, conversation modes, accepted corrections, and rewrite review help shape the draft. Only the corrections you approve become local learning data.
04
Output Lands in the Focused App
When the draft is ready, CoreScribe delivers finished text into the active app. Macros, workflows, and system tools stay available in Studio without changing the dictation path.
Learning boundary
Accepted corrections update your local profile only when you approve them.
CoreScribe uses accepted corrections, prompt drills, rewrite review, and local voice/profile context to improve behavior on this Mac. Style edits stay separate from learning, and routine use does not create a cloud training stream.
AcceptOnly approved corrections enter the profile.
SeparateStyle edits do not automatically become learning data.
LocalProfiles stay on the Mac by default.
DeliverFinished prose lands in the active app.
Private pilot
Bring CoreScribe into a real workflow.
Available by request for clinical, founder, and writing workflows. Apple Silicon optimized, macOS native, commercial-safe dependencies, and local-first by default.