Now open source · Learn more

Last Updated: October 11, 2026

Privacy Policy

HyperWhisper is built to keep your data on your device. You can transcribe entirely offline using local AI models, or choose from a range of cloud providers — including HyperWhisper Cloud, our built-in cloud service. Whichever option you choose, we never store your audio recordings or transcripts, with one exception for HyperWhisper Cloud audio. On the Chirp route, Cloud audio sits briefly in our own Google Cloud Storage bucket, is deleted after the request, and at most 1 day after a failure. The Chirp route is HyperWhisper Cloud's Google Chirp 3 model. A Chirp recording goes into the bucket when it is over 9.5 MB or when our server, estimating from its file size, judges it longer than about 55 seconds; other Chirp recordings are sent to Google without being written to it.

Ray Amjad LTD (Company Number 14506459, incorporated in the United Kingdom) is the provider of HyperWhisper and acts as the data controller for the limited personal data we handle.

What We Collect

  • Email: Used to deliver receipts, licenses, and support. We keep a record of the transactional emails we send you — the recipient address, email type, subject, send timestamp, and delivery status — retained for support, audit, and deliverability troubleshooting. We do not store the message body. When you request the download link from our website, we record the email address you submit so we can send it. We use the IP address of that request only to rate limit the form (10 requests per IP address per hour) and we do not store the IP address, browser user agent, or country with the email record. The rate limit does write the IP address to our rate-limit store, hosted by Upstash, in two forms: request counters, which expire within about two hours, and an hourly summary of requests per IP address and whether each one was allowed, which has no automatic deletion period today. Neither holds your email address, browser user agent, or country. Our transactional emails are delivered by Resend, which receives the recipient address and the message content in order to deliver them. For license key, Account Key, and sign-in link emails, that content includes the key or the sign-in link.
  • Order and billing info: Processed by our payment provider (Stripe) for purchases, refunds, and fraud prevention.
  • Crash and performance diagnostics: The HyperWhisper desktop apps for macOS, Windows, and Linux send diagnostics to Sentry, our error-monitoring provider. This is on by default. You can turn it off in Settings → General with the switch labelled "Error logging" on macOS and "Send error reports" on Windows and Linux. When you turn it off, the app stops collecting new diagnostics. On Windows and Linux, reports that are already waiting on your device can still be sent at that moment, and on macOS one final session record can be sent. The diagnostics include:
    • Crash and error reports with technical details about the error, and other diagnostic events and records about how the app is running. On macOS, this includes routine diagnostic log records that are sent even when nothing has gone wrong.
    • Release-health session data: records of app sessions and whether they ended in a crash.
    • Performance traces (timings of operations in the app) on macOS and Windows, and CPU profiles (samples of which parts of the app's code were running) on macOS. The Linux app does not currently record performance traces.
    • Identifiers: your IP address, the app version, your operating system version, your CPU architecture, and an identifier for your installation of the app that stays the same from one launch to the next and is linked to your sessions. On Windows and Linux, this also includes your computer's name, and on Linux the name of your user account on it.
    • Context about what the app was doing. For example, this can include the names of your audio input devices and the name of the app you were dictating into with the number of characters pasted (macOS and Windows), and file paths on your computer. On macOS, a file path can contain the name of your user account.
    No audio is attached to a report, so a report cannot give us your recordings. Report fields named for transcripts, text, or prompts are replaced with "[redacted]" before the report is sent. On macOS, an error report can also include recent lines from the app's own log. Those lines are filtered to remove some personal details, such as home-folder paths, email addresses, IPv4 addresses, and some text that looks like transcript content.
  • License information: The app contacts our licensing service to activate and validate your license. During this process, we receive your Account Key, a SHA-256 hash of your device's hardware identifier (we never receive the raw identifier), and your device's hostname. This information is used to enforce our fair usage policy. No audio or transcripts are transmitted. The licensing service also sees the IP address of each activation and validation request. We use it only to rate limit those endpoints (30 requests per IP address per minute). It goes into the same Upstash rate-limit store: its request counters expire within a few minutes, and the hourly summary of requests per IP address and whether each one was allowed has no automatic deletion period today.
  • Custom vocabulary: When you use HyperWhisper Cloud, our servers receive your custom vocabulary terms with each transcription or post-processing request and pass them to the provider that handles it.

We do not store your audio recordings or transcripts, with one exception. On the Chirp route, Cloud audio sits briefly in our own Google Cloud Storage bucket, is deleted after the request, and at most 1 day after a failure. The Chirp route is HyperWhisper Cloud's Google Chirp 3 model. A Chirp recording goes into the bucket when it is over 9.5 MB or when our server, estimating from its file size, judges it longer than about 55 seconds; other Chirp recordings are sent to Google without being written to it.

What Stays On Your Device

  • Recordings: Your audio stays on your device unless you export or share it.
  • Transcripts: Stored locally on your device; you control them.
  • Settings and vocabulary: App preferences and optional custom vocabulary are stored locally. When you use cloud transcription, the transcription provider receives your vocabulary terms with your audio as a spelling hint. With HyperWhisper Cloud, the terms pass through our servers to that provider, and to the post-processing provider when you use HyperWhisper Cloud post-processing. When you use local models for both transcription and post-processing, no transcription or post-processing provider receives your vocabulary. On macOS, if you turn on iCloud Sync for vocabulary, Apple stores your vocabulary (words and replacements) in your iCloud account.

Transcription Processing

HyperWhisper supports both local (on-device) and cloud-based transcription. When using local models, no audio data ever leaves your device.

When using cloud transcription, your audio is sent directly from your device to the cloud provider you choose. HyperWhisper does not proxy, store, or retain your audio or transcripts on our servers. HyperWhisper Cloud, below, has one exception. On the Chirp route, Cloud audio sits briefly in our own Google Cloud Storage bucket, is deleted after the request, and at most 1 day after a failure. The Chirp route is HyperWhisper Cloud's Google Chirp 3 model. A Chirp recording goes into the bucket when it is over 9.5 MB or when our server, estimating from its file size, judges it longer than about 55 seconds; other Chirp recordings are sent to Google without being written to it. The cloud provider you select may store or process your data according to their own privacy policies. You are responsible for reviewing those policies before use.

Third-party transcription providers (bring your own API key):

Third-party post-processing providers (bring your own API key) — these services, and any custom endpoint you add, receive your transcribed text and your custom vocabulary terms for correction and formatting:

HyperWhisper Cloud

HyperWhisper Cloud is our built-in cloud transcription service. It routes your audio through our edge servers to one of the providers listed below. We do not store your audio or transcripts — we only track credit usage — with one exception. On the Chirp route, Cloud audio sits briefly in our own Google Cloud Storage bucket, is deleted after the request, and at most 1 day after a failure. The Chirp route is HyperWhisper Cloud's Google Chirp 3 model. A Chirp recording goes into the bucket when it is over 9.5 MB or when our server, estimating from its file size, judges it longer than about 55 seconds; other Chirp recordings are sent to Google without being written to it.

Transcription (speech-to-text) providers:

  • Deepgram (Nova-3) — default provider. Also used as a fallback when Groq is unavailable in certain regions.
  • Groq (Whisper large-v3) — fastest option. Falls back to Deepgram if blocked in your region.
  • ElevenLabs (Scribe v2) — highest accuracy option.
  • Google (Chirp 3) — the Chirp route. Recordings over about 55 seconds or 9.5 MB pass through our own Google Cloud Storage bucket first, as described above.

Post-processing (text correction) providers:

  • Cerebras (GPT-OSS-120B) — default provider. Falls back to Groq on failure.
  • Groq (GPT-OSS-120B) — fallback provider. Falls back to Cerebras on failure.
  • Anthropic (Claude) (Claude Haiku 5.5). Falls back to Cerebras on failure.
  • SpaceXAI (Grok) (Grok 4.3). Falls back to Anthropic on failure.
  • OpenAI (GPT-5 mini or GPT-5 nano). Falls back to Anthropic on failure.
  • Google (Gemini) (Gemini 2.5 Flash, Gemini 2.5 Flash-Lite or Gemini 3.8 Flash). Falls back to Cerebras on failure.
  • Mistral (Mistral Small). Falls back to Groq on failure.

These providers receive your transcribed text, and the context described under "Screen Text and app context" below, to correct and format it. This post-processing request carries no audio. Grok and Mistral store request data for approximately 30 days for abuse monitoring, and do not use it to train their models. A request goes to the fallback named above in only two cases: the provider still returns a server error after any automatic retries, or its reply repeats our instructions instead of your corrected text. So one request can reach at most three providers in this list — for example Grok, then Anthropic after a server error, then Cerebras if Anthropic's reply repeats the instructions.

Where available, we have configured zero data retention on our provider accounts — for example, Deepgram's data retention is set to zero so that audio is deleted immediately after transcription. Tracking, voice data storage, and any optional data-sharing settings have been disabled on all provider accounts that offer those controls. However, we are not on enterprise plans with these providers, so their standard data processing terms still apply. Please review each provider's privacy policy linked above for full details on how they handle data.

HyperWhisper is fully open source under Apache-2.0 — including the Cloud backend. You can read the full source code for the Cloud backend and the apps on GitHub.

Medical Mode

With HyperWhisper Cloud, a mode that uses AssemblyAI can be set to the medical Transcription Domain. Our edge servers then send the audio to AssemblyAI with its Medical Mode add-on turned on. AssemblyAI processes it on HyperWhisper Cloud's own AssemblyAI account, under AssemblyAI's privacy policy, and we do not promise zero data retention there. A HyperWhisper Cloud mode can also use Deepgram's Nova-3 Medical or Nova-2 Medical model. These options tune transcription for clinical vocabulary. For these HyperWhisper Cloud medical requests, our edge servers do not write the audio or the transcript to our own storage. If Deepgram cannot take a request, HyperWhisper Cloud can send it to Groq or ElevenLabs as described above, and they do not use a medical model. Live streaming through HyperWhisper Cloud does not use a medical model. With your own AssemblyAI or Deepgram API key, you can instead choose AssemblyAI's Universal-2 Medical or Universal-3.5 Pro Medical model, or Deepgram's Nova-3 Medical or Nova-2 Medical model (Nova-3 Medical also for live streaming); your audio then goes from your device directly to that provider, under your own account, and not through our servers. If post-processing is on, the transcript also goes to the post-processing provider you use. Each provider processes your data under its own terms, and you must review them before you dictate about a patient.

Medical Mode is a vocabulary and formatting feature only. We are not a healthcare provider, and we do not sign a Business Associate Agreement (BAA). HyperWhisper, including HyperWhisper Cloud and Medical Mode, is not offered as HIPAA-compliant or as fit for regulated clinical record keeping.

Screen Text and app context

The post-processing request sent to the provider you choose (HyperWhisper Cloud, a bring-your-own-key provider, or a custom endpoint you add) carries more than your transcribed text. It includes your computer's name, local time, time zone and locale. It can also include the name of the app in front, the window or browser tab title, the website host, the type and label of the focused field, and up to 100 characters of the text in that field or of the text you have selected. A mode can also turn on Screen Text, which is off by default in every mode. When a recording starts, the app reads the text on the display that shows your active window (on macOS, the main display when it cannot find that window; on Linux, the area you select in the screenshot prompt) and adds up to 2,000 characters of it to the request. Text recognition runs on your device: the screenshot itself is never sent, only the recognized text. On macOS and Windows, no screenshot is taken when HyperWhisper itself is the app in front. When HyperWhisper identifies the app in front as sensitive, such as a password manager, the screen text and the field or selected text are left out, but the app name, the window or tab title, the website host and the field type and label are still sent. When post-processing is off, none of this is sent to a post-processing provider.

Payments and Licensing

  • Payments: All payments are handled by Stripe. We receive the minimum order metadata required to fulfill your purchase and provide support. Please review the Stripe Privacy Policy for details on how they handle payment data.
  • Past payments (before December 2025): Payments made before December 2025 were processed by Polar and their processors (e.g., Stripe). If you purchased through Polar, your order data may still be retained by them. See the Polar Privacy Policy for details.
  • Licensing: License activation/validation may send your Account Key and minimal device information to our licensing service to prevent abuse. No audio or transcripts are transmitted.

Cookies and Analytics

The app does not set tracking cookies. Our website uses strictly necessary cookies for authentication (when you log in to your account to manage your subscription or add credits), checkout, and license management. We do not use advertising cookies.

We use PostHog for analytics to understand how our website is used and to improve the experience. PostHog may set its own cookies. We do not use this data for advertising purposes.

The demo video on our home page is hosted by Bunny.net (Bunny Stream) and plays in an embedded player from iframe.mediadelivery.net. When the video player loads, Bunny.net receives your IP address, your browser user agent, and our website's address. The player may record plays and watch time, and it may store your playback settings, such as volume, in your browser's local storage. We do not send Bunny.net your email address, your account, your license (Account Key), your audio, or your transcripts through the player.

The desktop apps use Sentry, our error-monitoring provider, for crash and performance diagnostics, as described under What We Collect. You can turn it off in Settings → General.

Our website pages embed an AI support chat widget served by AgentStack (agentstack.build), another product by Ray Amjad. When the widget loads, that service receives your IP address, your browser user agent, and our website's address. Anything you type into the chat is sent to that service. The widget also loads its own analytics, a separate PostHog instance on AgentStack's domain (t.agentstack.build), which may set its own cookies or local storage. The widget has no access to your audio, your transcripts, or the desktop app.

Children's Privacy

HyperWhisper is not directed to children under 16. We do not knowingly collect personal information from children. If you believe a child has provided us information, contact us to delete it.

Your Rights

Because we do not store your audio or transcripts, requests to access or delete that data should be carried out on your device by you. The one exception is HyperWhisper Cloud audio. On the Chirp route, Cloud audio sits briefly in our own Google Cloud Storage bucket, is deleted after the request, and at most 1 day after a failure. The Chirp route is HyperWhisper Cloud's Google Chirp 3 model. A Chirp recording goes into the bucket when it is over 9.5 MB or when our server, estimating from its file size, judges it longer than about 55 seconds; other Chirp recordings are sent to Google without being written to it. For email, order, or license information we maintain, you may request access or deletion by contacting support. Some information must be retained for legal/accounting purposes.

If you dictated patient information with the Medical Mode options described above, through HyperWhisper Cloud or your own API key, we hold neither the audio nor the transcript, so delete them on your device as described above. Any copy a provider keeps is governed by that provider's terms.

International Transfers

Our payment and licensing providers, Upstash, and Resend may process limited personal information in multiple countries, including the United Kingdom, the European Economic Area, and the United States. Upstash hosts our rate-limit store. It also holds a one-hour cache, keyed by your Account Key, of whether the key is valid and its remaining credit balance. HyperWhisper Cloud checks request IP addresses against a block list held there. See the Upstash Privacy Policy for details. Where required, appropriate safeguards are used by those providers, including Upstash (for example, adequacy decisions, Standard Contractual Clauses, or equivalent mechanisms). The AgentStack support chat widget described above is served from the United States.

Changes to This Policy

We may update this policy. If we make material changes, we will update the date above and, where appropriate, notify recent purchasers.

Contact

Data Controller: Ray Amjad LTD (Company Number 14506459, United Kingdom). Questions or requests: hi@support.hyperwhisper.com