Updated September 7, 2026 · By Sumbat.T
Is OpenWhispr Safe? What Its Own Documents Say
The short answer
- Yes, and it is more verifiable than most. But the real answer is "which of the three modes are you running?" OpenWhispr is MIT-licensed and public, actively maintained (v1.9.2, 29 August 2026), and its own security page describes three processing modes with genuinely different privacy properties.
- In local mode, nothing leaves the machine. The privacy policy says no audio, transcription or agent prompt is transmitted to OpenWhispr or any third party. That is a structural guarantee, not a promise.
- The certifications are real, and better documented than most. An ISO/IEC 27001:2022 certificate numbered KCISM202607012, plus a SOC 2 Type 2 report for 1 May to 31 July 2026, available under NDA. Anyone telling you those two are decoration has not read the DPA. The HIPAA badge is the one to look at harder.
- One gap is worth your attention: the homepage and the DPA disagree about health data. The homepage badge says HIPAA "Compliant". Section 14 of the DPA says it "is not a Business Associate Agreement" and that PHI must not go through the cloud service without one.
- If you would rather the recording simply stayed on your disk without configuring anything, that is a product choice, not a settings choice. BlabbyAI writes its History audio to your own machine on Windows and never uploads it, at $8.49/month.
Dictation with nothing to configure
No models to download, no API keys to manage
BlabbyAI runs on Whisper v3 Turbo and types into any text field on Windows or in Chrome. On Windows, History saves every recording to your own disk. 60 credits a week free, no card.
"Is it safe" is four questions in a coat
Search for whether any dictation tool is safe and you will find confident answers pointing in opposite directions. That is not because anyone is lying. It is because the word covers four different questions, and OpenWhispr scores very differently on each. Separating them first makes the rest of this page much shorter.
1. Where does my audio go?
Entirely up to you here, which is unusual. Three modes, three different answers. This is the section that matters most.
2. Will it be used to train a model?
No, and it is committed to on the security page and again as a contractual term in the DPA. A strong answer.
3. Has anyone audited it?
Yes. A named ISO 27001 certificate and a completed SOC 2 Type 2 report, offered under NDA through a trust center.
4. Can I put regulated data through it?
This is where the marketing page and the legal document part ways. Read section 14 of the DPA before you decide.
The three modes, and why they are the whole answer
Most privacy write-ups treat a dictation app as one thing that either uploads your voice or does not. OpenWhispr does not work that way, and its own security page is explicit about it: the app converts speech to text on macOS, Windows and Linux, and you choose how the audio is processed. There are three modes, and they are not variations on a theme. They send your voice to three different places.
| Mode | Where the audio goes | Who can see it |
|---|---|---|
| Local models | Never leaves your machine | Nobody |
| Bring your own API key | Sent to your chosen provider | OpenAI, Deepgram, AssemblyAI, Groq or Mistral, on your account |
| OpenWhispr Cloud | Sent to their API, then discarded | OpenWhispr, plus OpenAI and Groq behind it |
Local models: the only mode with a structural guarantee
This is the mode people mean when they call OpenWhispr private. You download a Whisper or NVIDIA Parakeet model and transcription runs on your own processor. The privacy policy states the consequence without hedging.
"no audio, transcription, or agent prompt is transmitted to OpenWhispr or any third party", and the processing "happens entirely on your computer".
OpenWhispr privacy policy, last updated 12 August 2026
The reason this is worth more than an equivalent sentence from a closed-source vendor is that you can check it. The desktop app is MIT-licensed and public, and the security page makes the point itself: you can inspect exactly how audio is captured, processed and transmitted. There is no upload step to find in local mode, which is a different class of assurance from a company promising it deleted something.
The trade-off nobody mentions
Local models are large, and they ship inside the installer. On release 1.9.2 the Windows installer is about 252 MB, the Mac disk image about 332 MB, and the Linux AppImage 532 MB. That is the honest cost of never sending your voice anywhere: the model has to live on your machine, and your machine has to be fast enough to run it.
Bring your own key: OpenWhispr steps out of the path
In this mode you supply an API key for a provider you already have an account with, and, in the privacy policy's words, "audio is sent directly to that provider" under your own relationship with them. The listed options for transcription are OpenAI Whisper, Deepgram, AssemblyAI, Groq and Mistral, with OpenAI, Anthropic and Google Gemini available for the agent features.
The privacy question here is not really about OpenWhispr at all. Your audio is governed by whichever provider's terms you signed up to, and their retention policy is the one that applies. That is a reasonable design, and it is also the mode where people most often assume they are protected by OpenWhispr's promises when they are actually protected by OpenAI's.
OpenWhispr Cloud: real-time, then discarded
The managed option. Audio goes to their API, comes back as text, and is not kept.
"Audio is sent to our API, processed in real time, and discarded." Neither OpenWhispr nor its providers store the audio.
OpenWhispr security page
Behind that endpoint are OpenAI and Groq, with Parasail handling the agent and reasoning features, and the security page notes they may switch between providers while ensuring none of them trains on your data. The retention terms are published per vendor, which is more detail than most companies give: Groq is listed as opted out for their accounts, Parasail as zero data retention, and the memory and search vendors as no training on your data.
What leaves your machine, whichever mode you pick
Audio is only part of the picture, and this is where a lot of coverage stops too early. OpenWhispr publishes a data-at-a-glance table in its privacy policy, which is genuinely more transparency than the category norm. Here is what it says, condensed.
| Data | Where it lives | Notes |
|---|---|---|
| Voice and audio | Device only by default | Cloud subprocessors or your own provider, if you pick those modes |
| Transcriptions | Device SQLite | Not synced, so nobody else sees them |
| Account and billing | Neon and Stripe | Required for paid plans |
| Notes and agent chats | Device, or Neon if synced | Sync and agent are opt-in |
| Diagnostics | Off unless you enable it | No audio or transcription content included |
The one to note is the second row. Your transcriptions live in a local SQLite database and are not synced, which means the text of what you dictated stays with you even when you are using the cloud transcription mode. Notes and agent conversations are the exception, and only if you switch sync on, in which case they go to their cloud database, Neon, encrypted at rest.
Worth knowing, though it is about the website
The subprocessor list includes RB2B, described as B2B website-visitor identification for sales outreach, consent-gated and US visitors only, matching device and network identifiers to public professional profiles including name, company and work email. That runs on the marketing site, not inside the dictation app, and the distinction is important. It is still a striking thing to find listed by a privacy-first product, and you will not read it anywhere else.
The certifications are real. The homepage overstates one of them.
It would be easy, and wrong, to write this section as a takedown. When a young open-source project displays SOC 2 and ISO 27001 badges, the reasonable suspicion is that they are decoration. In this case they are not. The Data Processing Addendum names specifics that decoration does not come with.
"SOC 2 Type 2 report for the period 1 May 2026 to 31 July 2026, available under NDA; GDPR and HIPAA attestations issued via its compliance platform, 2026; ISO/IEC 27001:2022 certification, certificate KCISM202607012"
OpenWhispr DPA, section 10 (Audit), version 1.2
Elsewhere the DPA adds that the ISO certificate was issued by KVQA Certification Services, accredited by the Standards Council of Canada, with initial registration on 27 July 2026 and validity to 27 July 2027, and that the SOC 2 report covers Security, Confidentiality and Availability. There is a trust center at trust.openwhispr.com with a request-access path for the documents. That is a real compliance programme, and it is fair to say so.
Two observations stand, both factual. The SOC 2 audit firm is not disclosed publicly, only the report period, whereas the ISO certification body is named. And the security page makes no certification claims at all: the words HIPAA, SOC 2 and ISO 27001 do not appear on it. Every badge lives on the marketing homepage instead.
The one real gap: what the homepage says about health data, and what the DPA says
This is the finding that matters if you work with patient information, and it is the reason this page exists. The homepage carries a block headed "Compliance & certifications" with the subtitle "Everything your security team will ask for, already in place", and beneath it a HIPAA badge labelled "Compliant" reading "Whether you use local or cloud processing, we keep your health information safe."
Section 14 of the Data Processing Addendum says something considerably narrower.
"This DPA is not a Business Associate Agreement. Where the Customer is a Covered Entity or Business Associate under HIPAA, end users must not process protected health information through the cloud Services unless a separate Business Associate Agreement is in place between the parties."
OpenWhispr DPA, section 14 (HIPAA), version 1.2, updated 21 August 2026
A BAA is offered to customers on the Business plan and above, on request. So both statements are defensible on their own terms: the company does have a HIPAA attestation, and it does offer a BAA. But a clinician reading the homepage would reasonably conclude that cloud dictation of patient notes is covered, and the DPA says in plain words that it must not happen without a separate signed agreement they do not have by default. If that is your situation, the badge is not the document to read.
The general rule, not just here
On any dictation vendor, the marketing page and the DPA are written by different people for different reasons. When they disagree, the DPA is the one that governs what you are allowed to do. It takes five minutes to check and it is the single highest-value thing a security reviewer can do in this category.
What open source does and does not buy you
OpenWhispr is MIT-licensed, and the project is unmistakably alive: 7,276 stars, 928 forks, version 1.9.2 released 29 August 2026, and commits landing into September. It has a real security policy with private vulnerability reporting, a 48-hour acknowledgement target, and a scope list that names the right risks for an Electron app: remote code execution via crafted audio, privilege escalation through native key listeners and paste helpers, credential exposure, cross-site scripting in the renderer, insecure inter-process communication, and supply chain attacks through dependencies.
What it does buy you
The ability to verify rather than trust. In local mode there is no upload step in the code, and you or anyone else can confirm that. No policy page can offer the same thing.
What it does not
An audit. MIT carries no warranty, the repo has 333 open issues, and the cloud API is separate and closed. Readable is not the same as read.
A detail the reviews miss: most of its downloads are on Windows
OpenWhispr is often written up as a Mac tool, partly because the open-source dictation scene skews that way. Its own release assets say otherwise. Counting downloads on release 1.9.2 directly from the GitHub API, the Windows installer is the single most-downloaded file by a wide margin. Downloads are not users, and one release is not a history, but the gap is wide enough to be worth knowing.
| Platform | Downloads, release 1.9.2 | Installer size |
|---|---|---|
| Windows | 11,583 | ~252 MB |
| macOS (all four builds) | 6,725 | ~332 MB |
| Linux (deb, AppImage, rpm) | 2,658 | 426 to 532 MB |
That matters for the safety question in a practical way. The local mode, the one with the real privacy guarantee, is also the one that asks a Windows laptop to run a Whisper or Parakeet model well enough to keep up with your speech. On capable hardware that is fine. On an ordinary work laptop it is the reason people quietly switch to the cloud mode and stop getting the guarantee they chose the app for.
How the privacy model compares
Four tools, and the row that matters is where your audio physically goes. Prices are the current list prices at the time of writing.
| Tool | Transcription | Where the recording sits | Training | Price |
|---|---|---|---|---|
OpenWhispr | Local, BYOK or their cloud | Device only in local mode | No, contractually | Free, Pro paid |
 BlabbyAI | Cloud transcription | Local disk on Windows, never uploaded | Not used for training | $8.49/mo |
 Wispr Flow | Cloud only, no opt-out | Vendor servers, if cloud storage on | Opt-in, off by default | $15/mo |
 Superwhisper | Local models available | On device with local models | Not used with local models | $8.49/mo |
Read the table as three different trades rather than a ranking. OpenWhispr in local mode keeps audio off the network, and charges you model downloads, mode decisions and a machine fast enough to run them. Wispr Flow asks nothing of you and gives nothing back on this axis: transcription is cloud-only with no opt-out at any tier, at nearly twice the price. BlabbyAI is the one that gives you the recording without the setup, because on Windows the History feature writes every recording to your own disk the moment you stop speaking, before transcription even begins, and that audio is never uploaded. Nothing about your history touches a server, and there was nothing to configure to get that.
If you want the guarantee without the configuration
Everything above adds up to one trade. OpenWhispr hands you the controls, and the price of holding them is that you have to work them: choose a mode, download and manage models per machine, decide about sync, and read a DPA before any regulated data goes near it. That is a fair deal for people who want it.
Plenty of people want the outcome without the administration. BlabbyAI is built the other way round: there are no models to download, no engines to compare and no API keys to hold. You install it, press Ctrl+Space, and your words appear in whatever text field your cursor is in, whether that is Word, Outlook, Slack, Notion, an EHR text box or a code editor. It runs on Whisper v3 Turbo, supports 90+ languages with auto-detect, and costs $8.49/month for unlimited use, with 60 credits a week free and no card to start.
A BlabbyAI custom mode is a free-form instruction applied to what you just said. Source: BlabbyAI.
The feature worth knowing about is custom modes. A mode is a free-form instruction that runs over the words you just spoke, so you can have one that cleans up grammar, one that turns a rambling thought into an ordered set of bullet points, one that translates as you speak, or one that recognises your own spoken commands: tell it that whenever you say "new paragraph" it should insert a real break, and it will. It is closer to a programmable layer over your voice than a fixed feature list.
The Chrome extension is the other half of the product, and for a lot of people it is the half they use. Most dictation happens in a browser tab: Gmail, Google Docs, Notion, Jira, a CRM comment box, the prompt field of whichever AI tool you are talking to. The extension dictates into all of them with nothing to install, no 250 MB download and no account needed to try it, which also makes it the fastest way to find out whether dictation suits how you work at all.
How to decide, in one pass
You want the recording kept, and nothing to set up
BlabbyAI. History writes every recording to your own Windows disk with no models to manage, it types into any text field on Windows or in Chrome, and it is $8.49/month with 60 credits a week free.
You want to administer it yourself
OpenWhispr's local mode is built for that, and the cost is real: a model download per machine, a mode decision per user, and hardware quick enough to keep up with speech.
You handle regulated data
Read the DPA of whichever tool you are considering before the marketing page, and confirm in writing what agreement you actually hold.
Frequently asked questions
Is OpenWhispr safe to use?
For ordinary work, yes, and the architecture is more verifiable than most. The desktop app is MIT-licensed and public on GitHub, so anyone can read what it does with audio rather than trust a policy page, and it is actively maintained: version 1.9.2 shipped on 29 August 2026 against 7,276 stars. But "is OpenWhispr safe" has no single answer, because the app has three processing modes and they have genuinely different privacy properties. In local mode the privacy policy says no audio, transcription or agent prompt is transmitted to OpenWhispr or any third party. In bring-your-own-key mode your audio is sent directly to whichever provider you supplied a key for. In OpenWhispr Cloud mode audio is sent to their API, processed in real time and discarded. The honest answer is not yes or no, it is: which mode are you running? If you want a tool where the audio recording is written to your own disk by default and never uploaded, BlabbyAI does that on Windows for $8.49 a month.
Does OpenWhispr send my audio to the cloud?
Only if you choose a mode that does. The security page describes three modes and says you pick the one that fits your needs. With local models, the privacy policy states that no audio, transcription, or agent prompt is transmitted to OpenWhispr or any third party, and that the processing happens entirely on your computer. With bring-your-own-key, audio is sent directly to that provider under your own account and their privacy policy applies. With OpenWhispr Cloud, audio is sent to their API, processed in real time, and discarded. Two things do leave the device in more than one mode: account and billing data goes to their cloud database, and if you switch on the optional sync, notes and agent conversations are stored there too. Raw audio and transcriptions do not sync.
Does OpenWhispr have SOC 2 and ISO 27001?
It has a real ISO 27001 certificate and a completed SOC 2 Type 2 report, which is more than many tools in this category can show. The DPA names the specifics: an ISO/IEC 27001:2022 certificate numbered KCISM202607012, issued by KVQA Certification Services, accredited by the Standards Council of Canada, initially registered 27 July 2026 and valid to 27 July 2027; and a SOC 2 Type 2 report covering the period 1 May 2026 to 31 July 2026, covering Security, Confidentiality and Availability, available under NDA. There is also a trust center at trust.openwhispr.com with a request-access path. Two details are worth knowing. The SOC 2 audit firm is not disclosed publicly, only the report period. And the security page itself makes no certification claims at all: the badges live on the marketing homepage.
Is OpenWhispr HIPAA compliant?
This is the one place where the marketing page and the legal document say different things, and it matters if you handle patient data. The homepage carries a compliance block headed "Compliance & certifications" with the subtitle "Everything your security team will ask for, already in place", and a HIPAA badge marked "Compliant" reading "Whether you use local or cloud processing, we keep your health information safe." The Data Processing Addendum, section 14, is narrower: "This DPA is not a Business Associate Agreement. Where the Customer is a Covered Entity or Business Associate under HIPAA, end users must not process protected health information through the cloud Services unless a separate Business Associate Agreement is in place between the parties." OpenWhispr offers a BAA to customers on the Business plan and above, on request. So a reader who takes the homepage at face value could reasonably conclude they may dictate patient information through the cloud service, while the DPA says they must not without a separate signed agreement. Read the DPA, not the badge.
Does OpenWhispr train AI on my voice?
No, and it commits to this in more than one place. The security page says plainly that your data is not used to train any AI models, and that every provider used behind the cloud service is opted out of using your data for model training. The DPA repeats it as a contractual obligation: OpenWhispr will not use customer personal data to train, retrain or improve any artificial intelligence or machine learning model. In local mode the question does not arise, because nothing is transmitted at all. This is a genuinely strong position and one of the better answers in the category.
What does OpenWhispr actually store about me?
The privacy policy publishes a data-at-a-glance table, which is more transparency than most vendors offer. Audio is device-only by default. Transcriptions are stored in a local SQLite database on your machine and are not synced. Account and billing information goes to their cloud database, Neon, and to Stripe. Notes and agent conversations are device-only unless you switch on sync, in which case they go to Neon encrypted at rest. Diagnostic reporting is off by default and must be explicitly enabled, and the security page says no audio or transcription content is ever included in it. Passwords are hashed and never stored in plain text.
Which companies does OpenWhispr share data with?
The privacy policy names them in a subprocessor table, and the trust center lists thirteen. The core set is Neon for the account database and cloud sync, Stripe for payments, Resend for transactional email, Vercel for hosting and analytics, Google for OAuth and for Tag Manager, Analytics 4 and Ads, OpenAI and Groq for cloud speech-to-text behind OpenWhispr Cloud, and Parasail for agent inference. One entry is about the website rather than the app and is worth knowing about anyway: RB2B, described as B2B website-visitor identification for sales outreach, consent-gated and US visitors only, collecting device and network identifiers matched to public professional profiles including name, company and work email. That runs on the marketing site, not inside the dictation app, but it is a notable thing to find in the subprocessor list of a privacy-first product.
Is open source the same as safe?
It helps, but it is not a certification and it is worth being precise about what it buys you. OpenWhispr is MIT-licensed with 7,276 stars and 928 forks, so the desktop code genuinely can be inspected, and the security page points out you can verify exactly how audio is captured, processed and transmitted. Two limits. First, the cloud API service is separate and managed by them, so the part you can read is the client, not the server. Second, being readable is not the same as being read: the repo carries 333 open issues, and MIT means no warranty. Open source gives you the ability to check, which is real and rare. It does not give you an audit.
Is OpenWhispr a good choice on Windows?
It runs on Windows, and the download numbers suggest that is where most of its users are: on release 1.9.2 the Windows installer was the single most-downloaded asset at 11,583 downloads, against roughly 6,700 across all four Mac builds and 2,658 across the three Linux packages. The practical cost is size, because local models ship inside the installer: the Windows download is around 252 MB and the Linux AppImage is 532 MB. If you would rather not manage models, engines and API keys at all, BlabbyAI is a Windows-native app that runs on Whisper v3 Turbo with nothing to configure, and its History writes every recording to your own disk instead of a server.
What is the safest way to run OpenWhispr?
Pick local models during setup and do not switch on sync. That combination is the one where the privacy policy says nothing is transmitted to OpenWhispr or any third party, and it is a structural guarantee rather than a promise, because there is no upload step in the code path. You give up cloud speed, the agent features and cross-device access. If you do use the cloud mode, the things worth checking are that diagnostic reporting stays off, which is the default, and that you have read section 14 of the DPA before any regulated data goes anywhere near it.
The bottom line
OpenWhispr is safe, with a caveat that does most of the work: it is safe in the mode you selected, not as a product. It offers a mode in which your voice never leaves your computer, and it is open enough that you can verify that instead of believing it. The compliance paperwork behind it is real too, with a named ISO certificate and a completed SOC 2 report. Anyone calling those two badges decoration has not read the DPA.
The two things to carry away are these. First, the guarantee lives in the mode, not in the brand, and it quietly disappears the moment you switch to cloud transcription for speed. Second, if you handle health data, the homepage and section 14 of the DPA do not say the same thing, and the DPA is the one that governs. Check which one you were relying on.
And if reading this has mostly convinced you that you do not want to think about processing modes at all, that is a legitimate conclusion. BlabbyAI exists for that: install it, press a key, talk, and get clean text in any application, with your recordings kept on your own machine and nothing to configure. It is free to start and $8.49/month if you keep it.
Dictation without the configuration
No models to download, no engines to compare, no API keys to hold. Press a key, talk, and clean text lands in whatever you are typing into. On Windows your recordings stay on your own disk. 60 credits a week free, no card.
Sources
- OpenWhispr, Local Speech-to-Text Privacy & Security. The three processing modes, audio handling and training commitments.
- OpenWhispr, Privacy Policy, last updated 12 August 2026. Data-at-a-glance table and the subprocessor list.
- OpenWhispr, Data Processing Addendum, version 1.2, updated 21 August 2026. Section 10 (Audit) and section 14 (HIPAA).
- GitHub, OpenWhispr/openwhispr and its security policy. Licence, release 1.9.2, per-asset download counts and vulnerability scope, read from the GitHub API on 7 September 2026.
- BlabbyAI, dictation software and Windows speech to text.