Legal

Privacy Policy

Last updated: 16 July 2026

TalkToLaTeX turns spoken mathematics into LaTeX. This policy explains what we collect, how your voice is processed, who processes it, and the choices you have. We've kept it specific and plain. No boilerplate that hides what actually happens to your audio.

1. Who we are

TalkToLaTeX ("we", "us") provides a desktop app, a browser extension, and a website that convert dictated speech into LaTeX. For privacy questions you can reach us at privacy@talktolatex.com.

2. Information we collect

  • Account information: your email address and a securely hashed password (we never store your password in plain text). If you sign in with Google, we receive your verified email and a Google account identifier instead of a password.
  • Voice and dictation content: the audio you record while dictating, the transcript produced from it, and the LaTeX generated from that transcript. How long these exist is described in section 4.
  • Document content: none. We do not read your document. The extension converts only the words you dictate, and nothing from the document you are dictating into is sent anywhere. This is described in section 4.
  • Usage data: counts of conversions, which speed/quality tier was used, and timestamps, so we can meter usage and keep the service running.
  • Technical data: your IP address (used transiently for rate-limiting and abuse prevention), and a sign-in token plus minimal preferences stored in your browser's local storage. The "try it out" demo also uses Cloudflare Turnstile to tell humans from bots.

3. How we use your information

  • To provide the core service: transcribe your speech and convert it to LaTeX.
  • To create and secure your account, and to sync your dictation history across devices if you are signed in.
  • To meter usage, enforce rate limits and our spending cap, and prevent abuse.
  • To send essential account emails (for example, email verification and password resets).

We do not sell your personal information, and we do not use your voice or transcripts to advertise to you.

4. Voice data & AI processing

This is the part that matters most, so here is exactly what happens when you dictate:

  • Audio is captured only while a dictation you started is running — from the moment you start it until you stop it (or a 5-minute safety cap ends it). Nothing is recorded in the background.
  • The audio is sent over an encrypted (TLS) connection to our backend, which forwards it for speech-to-text to OpenAI and, in the browser extension, also to Groq. The extension sends the clip to both and uses whichever answers first, because that race is what makes dictation feel instant. Both therefore receive your audio.
  • The transcript is then converted into LaTeX. In the browser extension this is done by Google (Gemini), with DeepSeek as an automatic fallback when Gemini fails or returns nothing usable. In the macOS app it is done by Anthropic (Claude). The result is returned to you.
  • We do not send your document. Only the transcript of what you just said leaves your browser. Until 16 July 2026 the extension also sent the conversion provider a window of text around your cursor, plus your preamble lines and \label keys, so the output could match your own notation. That was removed: no part of your document — published or unpublished — is sent anywhere. One exception, stated plainly: if your Overleaf project contains a .bib bibliography file, its entries (citation key, author, title and year) are sent with your dictations, so that a dictated citation resolves to the right entry instead of a fabricated one. That is bibliography metadata, not your writing, and it is the only thing from your project that leaves your browser.
  • We do not store your audio. On the macOS app it is held only in memory for the moment it takes to transcribe. We do not retain the audio after the response is returned.
  • These providers process the data as our service providers under their API terms, and their terms are not identical, so we will not flatten them into one reassuring sentence:
    • OpenAI and Anthropic: under their API terms your content is not used to train their models, and is retained only briefly for abuse monitoring before deletion.
    • Groq: states it does not use API inputs or outputs to train or fine-tune models, and does not retain inference data by default, aside from temporary logs when investigating errors or abuse.
    • Google (Gemini): we use Gemini on paid terms, which matters and is deliberate. Under those terms Google states it does not use your prompts or responses to improve its products, and logs them only for a limited period to detect abuse and keep the service secure. Google's free tier says the opposite (content used to improve Google's products, and human reviewers may read it), which is why we do not use it.
    • DeepSeek: its platform terms state it may use inputs and outputs, de-identified, to a minimal extent to maintain and improve its services and underlying technology. DeepSeek is based in China, so this processing takes place outside the EEA.
  • The website demo additionally keeps an anonymous per-IP counter (just a number of free tries used that day) so we can enforce the daily limit. It is not linked to your identity and expires automatically.

5. Storage and retention

  • Dictation history. The macOS app stores your dictation history locally on your own device. The browser extension stores it locally too, and while you are signed in it also saves each entry (your transcript and the LaTeX produced from it) to our systems so your history is available on your other machines. This happens on every dictation while signed in; there is currently no switch to turn it off, and if that matters to you, dictate signed out or delete your account to remove it. It does not include your audio.
  • Account data is retained for as long as your account exists. When you delete your account from within the app, we delete your account record and associated stored history.
  • Operational data such as usage counters and abuse-prevention records is kept only as long as needed for those purposes and then expires.

6. Third-party processors

We rely on a small set of providers to run the service. They process data on our behalf:

  • OpenAI: speech-to-text transcription.
  • Groq: speech-to-text transcription (browser extension, raced against OpenAI for speed).
  • Google: converting transcripts into LaTeX in the browser extension. It receives your transcript (and, if your project has one, the bibliography entries described in section 4) — not your document. Separately, and only if you choose it, "Continue with Google" sign-in.
  • DeepSeek: converting transcripts into LaTeX in the browser extension when the primary provider fails or returns nothing usable. Processing takes place in China.
  • Anthropic: converting transcripts into LaTeX in the macOS app.
  • Cloudflare: hosting, network security, and the Turnstile human-verification on the demo.
  • Resend: sending essential account emails.

7. Cookies and local storage

We do not use advertising or tracking cookies. We use your browser's local storage to keep you signed in and to remember basic preferences, and Cloudflare Turnstile may set a token to verify you are human on the demo. Clearing your browser storage will sign you out.

8. Security

Connections use TLS encryption. Passwords are salted and hashed; our API keys for the providers named in section 6 are held only on our backend and are never exposed to the app or the browser. No system is perfectly secure, but we design to keep your credentials and content protected.

9. Your rights

You can access the email on your account, delete your account and stored history from within the app, and contact us to ask what we hold about you. Depending on where you live (for example, under the GDPR), you may also have rights to access, correct, export, or erase your personal data. Email us to exercise them.

10. Children

TalkToLaTeX is not directed to children under 13 (or the minimum age in your country), and we do not knowingly collect their personal information.

11. Changes to this policy

We may update this policy as the product evolves. We will revise the "last updated" date above, and significant changes will be highlighted in the app or on this page.

12. Contact

Questions about privacy? Email privacy@talktolatex.com.

Note: TalkToLaTeX is in active beta. This policy describes our current practices in good faith; it is provided for transparency and is not legal advice. If you operate this service, have it reviewed by a qualified professional for your jurisdiction.