Nobody enters medicine because they want to type. The training covers anatomy, pharmacology, differential diagnosis, and how to sit with a patient who has just received bad news.
What it does not cover is how to spend two hours after a full clinic day turning encounters into notes that satisfy billing, legal, and continuity-of-care requirements all at once. That work has grown steadily for two decades, and for most clinicians it now runs longer than the patient contact that generated it.
Where the hours actually go
Clinical documentation is one of the biggest time sinks in medicine, and the reason has less to do with typing speed than with what a note has to accomplish. A single encounter note carries the patient’s account, the exam findings, the reasoning behind a diagnosis, and the plan going forward, and each of those pieces has to be complete enough to defend later. Medical dictation was the original answer to this, letting a clinician speak the note instead of typing it. The catch with most medical dictation tools is that they stream audio to remote servers, which means diagnoses, medications, and history leave the workstation before they become text.
Clinicians who want to close that gap can run OpenWhispr for medical transcription directly on the workstation, so patient audio is never sent out for processing. Core dictation keeps working with the internet disconnected, which also removes the outage risk that comes with hosted tools.
The privacy question practices actually have to answer
Any tool that touches patient audio creates a chain-of-custody question. Once that audio reaches a vendor’s servers, the practice inherits a set of obligations: what the vendor retains, for how long, under what agreement, and what happens if the vendor is breached. Those are answerable questions, but answering them takes legal review and ongoing vendor management that small practices are rarely staffed to handle.
Local processing changes the shape of the problem rather than the amount of diligence required. When audio never leaves the machine, there is no vendor retention policy to audit and no per-minute processing fee tied to how much a clinician speaks. What still needs review is the full workflow, including anything the practice adds on afterward, and no vendor can settle that on a practice’s behalf.
Getting text into the record
The gap between spoken words and a completed chart entry is where most dictation workflows break. Some tools require a custom integration with the electronic health record, which means waiting on IT, vendor coordination, and whatever the integration costs. Others produce a transcript in a separate window that the clinician then has to copy and paste, which reintroduces the manual work the tool was supposed to eliminate.
The alternative approach is a system-wide hotkey that types into whatever field currently has focus. That works across record systems, browser-based charts, standalone charting tools, word processors, and email without anyone building anything. It also means the clinician learns one behavior and applies it everywhere rather than learning a different workflow for each system in the building.
Accuracy the clinician controls
Speech recognition trained on general language handles clinical dictation well across the range of material a practice produces daily. Drug names, uncommon abbreviations, and specialty jargon are the places where any recognition system has to work hardest, and the difference lies in whether the clinician has any control over how the system handles them.
Model size is adjustable, which is where that control comes from. Larger local models handle dense clinical language better, at the cost of some speed, while smaller models suit quick notes where turnaround matters more than capturing an unusual term perfectly. A clinician writing detailed consult letters and one working through brief follow-up notes have different needs, and being able to change that setting rather than accept a fixed default means the tool adapts to the specialty instead of the specialty adapting to the tool.
Different documentation, different demands
Primary care notes tend to be written immediately after a visit while the details are still sharp, which favors a workflow with almost no friction between finishing with the patient and starting the note. Waiting until the end of the day means reconstructing from memory, and reconstruction is where errors and omissions enter the record.
Therapy and behavioral health notes carry different sensitivity. Session content is often the most private material a practice holds, and progress notes follow structures that differ meaningfully from standard medical format. Radiology-style reporting and specialty consults sit somewhere else again, with findings, impressions, and procedure detail going into reporting fields that may have nothing to do with a conventional chart. Any tool covering all three has to be indifferent to where the text lands.
Why open code matters to a security team
Most practices evaluate software by reading marketing material and asking the vendor questions. That works until someone in the organization needs to verify a claim rather than accept it. When the codebase is public, an IT security lead, privacy officer, or outside auditor can inspect what the software actually does and confirm that processing happens where the vendor says it does.
That verification path does not exist with closed products. The vendor’s word, backed by whatever contractual language the practice negotiates, is the whole assurance. For organizations with real security review processes, being able to hand an auditor the source rather than a datasheet shortens the evaluation considerably and removes the uncomfortable position of trusting a claim nobody outside the vendor can check.
What changes once processing moves onto the workstation
The architecture is the whole argument. When speech-to-text runs on the machine in front of the clinician, patient audio never enters a transmission chain, transcription history stays on local storage, and the software keeps working when the connection drops. There is no per-minute cost tied to how much a clinician speaks, which removes the quiet pressure to keep notes short that metered pricing creates.
Open code closes the loop. A practice does not have to take a vendor’s description on faith when its own security lead, privacy officer, or outside auditor can read the source and confirm exactly where processing happens. That combination of local execution and verifiable code gives a practice something a service agreement cannot: proof rather than a promise.
