Legal dictation software that keeps client information on your machine

Whisperstream is a $29 one-time dictation app for Windows. In Lockdown Mode, transcription, AI cleanup, and transcript storage all run on your own PC, and cloud providers are turned off. That locality is the whole argument of this page: when nothing you dictate leaves your machine, no vendor receives client information for this workflow, and the third-party questions your confidentiality analysis would otherwise have to clear fall away for this data path. This page is not legal advice; it gives you the facts your own analysis needs.

The short answer

No dictation software is approved or certified by any bar or by the ABA, and no product can satisfy your confidentiality duties for you. What you can evaluate is where client audio goes, who can access the text, and what safeguards a tool supports.

The confidentiality angle

ABA Model Rule 1.6(c) requires reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client. Dictating a client matter into software is exactly the kind of handling those efforts cover.

Why locality matters

Cloud dictation sends client audio to a vendor, a disclosure surface you must vet under the reasonable-efforts factors, and ABA Formal Opinion 512 requires lawyers to evaluate disclosure risks before putting client information into a generative AI tool. Local-only dictation removes that surface for this workflow.

Who this is for

Solo and small-firm lawyers dictating on a Windows PC without an IT department or a document management admin to run vendor reviews. Dictating PHI in a healthcare practice instead? See HIPAA dictation software.

Start with the rule rather than any product page. ABA Model Rule 1.6(a) provides that a lawyer shall not reveal information relating to the representation of a client unless the client gives informed consent, subject to limited exceptions. Rule 1.6(c) adds the duty that matters for tooling: "A lawyer shall make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client." Dictating a client matter into software is handling information relating to a representation, so the choice of dictation tool sits squarely inside that duty.

Reasonable efforts is a flexible standard, and Comment [18] to Rule 1.6 names the factors that inform it: "the sensitivity of the information, the likelihood of disclosure if additional safeguards are not employed, the cost of employing additional safeguards, the difficulty of implementing the safeguards, and the extent to which the safeguards adversely affect the lawyer's ability to represent clients". In plain terms: how sensitive is what you dictate, how likely is it to get out through the tool you chose, and what would guarding it cost you in money and workflow.

Generative AI raised the stakes. ABA Formal Opinion 512 (July 29, 2024) says that "before lawyers input information relating to the representation of a client into a GAI tool, they must evaluate the risks that the information will be disclosed to or accessed by others outside the firm", and it adds that because many self-learning GAI tools are designed so their output could lead directly or indirectly to disclosure of client information, a client's informed consent is required before inputting information relating to the representation into such a tool. Earlier, Formal Opinion 477 (May 11, 2017) had already applied the reasonable-efforts lens to transmitting client information over the internet. None of this page is legal advice or an ethics opinion; it is the vocabulary your own analysis uses, and the rest of the page maps one tool to it.

Run the reasonable-efforts factors against cloud dictation and a familiar checklist appears. Client audio leaves your machine for a vendor's servers, so the vendor becomes a disclosure surface you have to vet: what its terms say, whether recordings or transcripts are used to train models, how long anything is retained, who inside the vendor can access it, and what happens in a breach. Opinion 512 adds the generative AI question: before client information goes into a GAI tool, you must evaluate the risks of disclosure to others outside the firm, and for many self-learning tools informed client consent is required first. All of that is doable, and plenty of firms do it. It is also real work, and it has to be redone whenever the vendor's terms change.

Local-only dictation removes that surface for this workflow. In Lockdown Mode, transcription runs on your own PC, AI cleanup uses an on-device model, and cloud providers are turned off. Your audio, transcripts, and dictated text never leave your machine, and Lanreal, the company behind Whisperstream, never receives them. No vendor receives client information for this workflow, so for this data path there are no vendor terms to parse, no training or retention policy to clear, and no third-party generative AI input for Opinion 512's consent question to attach to.

Simplifies is the honest verb, not solves. Reasonable efforts still cover things no app controls: who can physically reach the PC, how your Windows account is shared, what happens to the text after it is pasted into an email or a filing, and every other tool in your stack that does send client information somewhere. The analysis and the determination remain yours, made for your own practice. What a local-only tool changes is the size of the question: for dictation itself, the third-party column is empty.

Free local dictation tools clear the transmission question cleanly: no client audio reaches a vendor, so there is no third party to vet for this data path, just as in Lockdown Mode. But Rule 1.6(c) speaks of unauthorized disclosure and unauthorized access, and the access half lives on the machine. Per their published source code, the popular free local dictation tools we audited in July 2026 store plaintext transcripts and unencrypted audio of every dictation, saved unconditionally, with no history-off switch and no lock.

Run that through the Comment [18] factors and the gap is plain: months of dictated client matters, readable to anyone who can reach the PC, its backups, or a shared sync folder, is precisely the unauthorized-access surface the reasonable-efforts standard asks you to weigh. Free and local answers the transmission question; it leaves the access question open. We published the full tool-by-tool findings in our dictation app privacy audit.

The Model Rules do not come with a product checklist, but your reasonable-efforts analysis needs concrete facts about what a tool does. The mapping below states what Lockdown Mode does in the duties' own vocabulary so you can drop it into that analysis. It is a description of software behavior, not an ethics opinion and not legal advice, and the reasonable-efforts determination is yours to make for your own practice.

ConfidentialityModel Rule 1.6(c)

In Lockdown Mode your audio, transcripts, and dictated text are never transmitted off your device, transcript history is encrypted at rest, and cloud providers are turned off. This is designed to support the reasonable efforts Rule 1.6(c) asks for; whether your efforts are reasonable is your determination, not the software's.

Access controlThe shared-office reality

Your transcript history is locked behind a password. A key derived from that password wraps the encryption key. Without your password or your one-time recovery code, the history stays encrypted. Transcripts lock again automatically after a period of inactivity that you set.

Client files at restEncryption

Transcript history is encrypted at rest with AES-256-GCM. The authenticated encryption is designed to detect tampering, which helps confirm your transcripts have not been altered.

Audit trailContent-free by design

A private, content-free access log records when your history is opened, locked, and unlocked, and when an unlock fails. It never records what your transcripts say, and it stays on your device.

Generative AIABA Formal Opinion 512

AI cleanup in Lockdown Mode runs on an on-device model, and cloud providers are turned off, so no client information is input into a third-party generative AI service in this workflow. That simplifies the disclosure-risk evaluation Opinion 512 requires: for this data path there is no outside recipient to evaluate. The evaluation itself, and your other tools, remain yours.

The determination

Whether your safeguards amount to reasonable efforts under Rule 1.6(c) is a judgment only you can make for your own practice, weighing the factors in Comment [18]. No software makes it for you.

Physical and office security

Who can physically reach the PC that holds your transcript history, which screens are visible to clients or opposing parties, shared printers, and everything else about the room stays yours to manage.

User identity

Whisperstream relies on your Windows account to identify the user. The Lockdown password is a second gate on the transcript history, not a separate identity for each person who shares the device. If several people share one Windows login, they share one identity.

Recovery

If you lose both your password and your one-time recovery code, your transcript history cannot be recovered. That is by design: the encryption key is wrapped by a key derived from your password, so there is no reset path through Lanreal.