Who Trains on Your Audio

Your voice usually leaves your machine. Nine of the fourteen tools below send audio to a server every time you use them, and two more fall back to a server when your hardware or your language is not supported.

By the VoiceBoard editorial teamUpdated August 28, 2026

How to read this page

Every product got the same three questions.

1. Does the audio leave the device?
2. Is the vendor training on it by default?
3. How do you turn that off, and which document says so?

For the third one we care about where the answer lives, because that turns out to decide how much it is worth. A privacy policy or a terms of service carries an effective date, a version history and legal weight. A page headed "Privacy and Security", with a shield icon and no date, carries none of that. Help centre articles sit in between: usually accurate, usually current, and rewritten without notice.

So each row below says which kind of text the answer came from:

• contract: privacy policy, terms of service or data processing agreement, with an effective date
• help: the vendor's own help centre or documentation, usually dated only in relative terms
• shop window: a marketing page with no date on it
• architecture: nothing is promised because nothing is possible, the audio never reaches the vendor
• not stated: no claim anywhere, which is not the same as a promise

We read all of these on the dates shown. Policies change, so this page is re-checked on a fixed schedule and after any vendor announcement, and the date it was last read sits at the top of the page.

The table

Sixteen rows for fourteen products, because local Whisper and Handy are the same answer arrived at two ways and both belong in a table about where audio goes.

Where the shop window and the contract disagree

This is the part nobody else publishes, so here are the pairs, with both quotes.

Rev

The security page says: "we'll never train external LLMs on your data". The help centre goes further and says Rev "does not use customer data, including transcripts, captions, or any other uploaded content, to train AI models or large language models (LLMs)", and that "Rev only trains proprietary automatic speech recognition (ASR) models".

The terms of service, updated 15 May 2026, say: "your Customer Content will be analyzed by our ASR models and other Rev artificial intelligence models and may be used for continuous training of those models".

The help centre narrows it to speech recognition. The contract keeps the door open for the rest. If those two ever have to be reconciled, the one with the effective date wins.

There is a second thing in Rev's help centre worth knowing, because it is unusually candid: "Rev now uses data perpetually, not just while being an active customer". Cancelling your account does not withdraw what you already uploaded.

The opt-out is real and easy to miss, because it is not a setting: "Customers are able to opt out of sharing their data for training purposes at any time, by emailing support@rev.com". That article was last updated on 12 September 2025. Rev's own help pages block our requests, so we read it in an archived copy from 23 October 2025 and flag that rather than pretend otherwise.

One more gap on the same product. "Every transcriptionist signs a strict NDA" and identities are "thoroughly verified" are claims from the security page. The contract says only that Rev "will enter into a written agreement with each subcontractor". The vetting is probably exactly as described. It is simply not a promise you could hold anyone to.

Otter

The security page says: "No customer data will be used to train or improve our AI Service Provider(s)' artificial intelligence models/algorithms."

Read that twice. It is a promise about other companies. Two paragraphs down the same page says what Otter does itself: "Otter uses a proprietary method to de-identify user data before training our models so that an individual user cannot be identified."

The privacy policy, effective 16 June 2026, is blunter: Otter processes your data to "Improve and monitor the Services, including training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)". The same policy lists "Data labeling service providers who provide annotation services and use the data we share to create training and evaluation data". The terms of service, effective 19 September 2025, lock it in: Otter "may collect, create, process, transmit, store, use, and disclose aggregated and/or deidentified data derived from Data or use of the Services for its business purposes, including for machine learning and training".

Two things follow that you will not find in the roundups.

There is no training opt-out for an ordinary account. We looked in the privacy policy, the terms, the security page and the help centre. The policy offers an opt-out from the "sale" of personal information and from targeted advertising, nothing more. The one help article with "opt out" in the title limits itself in its first line: "Submitting a form below is only for opting out of AI features, including Otter Chat & Topics. This does not affect any of the other features or integrations." For organisations there is a route, and it is a conversation rather than a switch: Otter's enterprise page says "Keep your organization's data out of AI model training. Contact your account manager to get started."

The widely repeated claim that you can opt out in account settings is not supported by any Otter document we could open. It appears in AI answers and in third-party trust indexes. It is not in the policy. If you have read somewhere that you already turned this off, check what you actually turned off.

And "de-identified" is carrying a lot of weight in that sentence. Stripping the account details from an audio file does not strip the voice, the names spoken inside it, or the substance of what was said.

Descript

The security page: "This option is disabled by default and can only be enabled by you."

The help centre, describing how to switch it off: "Scroll down to the Share data with Descript section and click Allowed to toggle off data sharing." It also notes that "Enterprise drives do not have the option to toggle data sharing, which is disabled by default", a clarification that makes no sense if it were already off for everyone.

Both texts belong to Descript. One says the toggle starts off. The other describes it reading "Allowed" and tells you to click it. We cannot resolve that without creating an account, so we are telling you what each page says and where the switch is: App Settings (Command or Control plus comma), the Profile tab, the "Share data with Descript" section.

Credit where it is due, because Descript's help centre also contains the clearest statements on this page: "Current AI models in production use no Descript user data", and "We have no plans to use any user's data who has opted out of data sharing at any stage of research, development, or production". The privacy policy, last updated 14 April 2025, backs the right up: "You can opt out of having your Projects used to improve the Descript Service by disabling the Share Data with Descript setting."

Voice cloning is a separate matter and Descript says so plainly: there, "we use the audio that you shared as 'Training Audio' to improve our service", and "Descript employees may listen to samples".

superwhisper and MacWhisper

Both sell privacy honestly and both have a privacy policy that has fallen behind the product.

superwhisper's terms, updated June 2026, say what you would want: "your data is not used to train, fine-tune, or improve AI models", and audio "is not retained on our servers". Its privacy policy, last updated 19 June 2024, describes an app "that transcribes all audio data locally on your device" and concludes "there is no need for data transmission". The current app offers cloud voice models and cloud language models from Deepgram, Anthropic, OpenAI and Groq. The commitment you would rely on lives in the terms and the documentation, not in the document labelled privacy policy.

MacWhisper's privacy policy, last updated 14 February 2024, says: "MacWhisper does all it's functionality on your device. No data (audio, text or other) leaves your device." The pricing page on the same site lists eight cloud providers you can connect with your own API keys, plus workflow steps that upload finished transcripts to Notion, Zapier or a webhook. Transcription itself really is local, and that is the reason to use it. The policy simply describes a version of the app from two and a half years ago and has nothing to say about the cloud half.

Neither is a scandal. Both are a reason to check the date on the document before you quote it at your security team.

The tools that keep the audio

Nothing leaves the machine

VoiceInk is the cleanest answer in the table, and unusually its contract is stricter than its marketing. The privacy policy, last updated 20 April 2026, says: "By default, all transcription processing happens entirely on your device using local AI models. No data leaves your computer unless you explicitly choose to enable optional cloud services." It then lists the exceptions itself, which almost nobody does: cloud transcription sends the audio file to the provider you picked, cloud enhancement sends only the text, and clipboard or window context goes only if you switch those on. The vendor cannot train on you either way: "VoiceInk does not operate cloud servers for storing user data." The app is GPL-3.0 on GitHub, so the claim is inspectable rather than merely stated.

One thing the marketing does not tell you and the policy does: your own transcripts and audio pile up locally forever unless you configure cleanup. "Transcriptions: Kept indefinitely by default until you delete them." If the reason you went local is a laptop that travels, turn the automatic deletion on.

Local Whisper, whether through Buzz, whisper.cpp, WhisperX or a Mac app, makes no promises for the same reason: there is no vendor in the middle. That is a stronger guarantee than any sentence a lawyer can write, and it has one honest limit. Local processing protects the route, not the file. The recording still sits on a disk that might be shared, backed up or unencrypted.

Handy is free, MIT licensed and, in its own words, "works completely offline". It has no privacy policy at all. On a product whose source you can read, that bothers us less than it would elsewhere, and we note it rather than smoothing it over.

superwhisper belongs here whenever you pick a local voice model and skip the language model step. Its documentation spells out the configuration: use Voice Mode, "which outputs raw transcribed text with no language model involved. This is the simplest fully-local configuration." Pick a cloud model instead and you are in the other half of the table, with a zero retention commitment covering it. Note the vendor's own caveat, which is a good sentence and a rare one: zero retention "applies to API usage only", so dictating into the ChatGPT app puts your words under OpenAI's consumer terms rather than superwhisper's.

MacWhisper transcribes locally, costs €64 once with lifetime updates, and turns into a cloud product the moment you paste in an API key. Both halves are fine. Know which one you are using.

Cloud, no training

Voicy is a cloud product with no offline mode, and it says so in its own comparison table. Audio goes to Voicy's servers and from there to Groq, which runs the Whisper model. What makes its policy unusual is that it explains the mechanism instead of just promising an outcome: "We do not store any recording or transcription data. The Groq API does not retain any information we send it as we have enabled their Zero Data Retention settings and we do not use Groq's batch or fine-tuning endpoints." Naming the specific endpoints they do not use is more informative than most of the reassurance on this page.

Two notes for accuracy. The sentence "We do not use your recordings to train an AI model or for any other purpose" is on the site's FAQ, not in the privacy policy, and the privacy policy carries no date at all. And two Voicy documents describe different geography: the privacy policy says servers are "hosted in the EU and US", while the security policy, version 1.3 dated 31 July 2025, says "All processing occurs in USA-based infrastructure". If EU processing is a requirement for you, ask before you buy rather than picking whichever page you found first.

Transkriptor states the no-training answer with no conditions attached to it: "We do not use your content to train, develop, or improve any AI or machine learning models", and the same sentence continues, "not for our own benefit, and not for any third party. Your transcriptions remain exclusively yours." Around that it puts the things that make a promise checkable: processing "hosted exclusively within the European Union", ISO 27001 and SOC 2, a deletion window you set yourself "from as short as 1 minute to up to 1 year", and a right to demand deletion within ten business days with a certificate to prove it.

Where that sentence lives is worth knowing, and it is the same yardstick we applied to everyone else. It is on the security page and in the Pro feature list. The privacy policy, terms of service and data processing agreement are published as a single page that carries no effective date, and the words train, machine learning and model do not appear in it. The closest legal support is the purpose limitation in the DPA: the processor "shall not process Client Personal Data for any purposes other than those specified in this Agreement". That is a working protection and an indirect one. On the pricing page the no-training line is listed under Pro, and it is not listed under Lite or Team.

Notta says nothing about training anywhere. We read its privacy policy and terms, both effective 11 March 2025, its security page and its help centre. The only line about improvement is about behaviour rather than content: "We may use the information gathered to perform statistical analysis of user behavior or to evaluate and improve the Notta Service." The one flat denial of training applies only to Google Workspace data and is boilerplate Google requires.

Silence is not a promise. It is also not an accusation, and there is one detail in Notta's favour that we have not seen mentioned anywhere: its content licence is the narrowest of the five transcription services here. You grant Notta rights "only as reasonably necessary" to provide, maintain and update the service, fix problems, comply with law, or as you permit in writing. Training is not on that list, so the contract does not appear to authorise it. That is an argument, not a guarantee.

Claims that Notta trains on free-tier data, or deletes audio after ninety days, circulate in reviews. Neither is in any Notta document we could find.

Cloud, training on by default

Wispr Flow processes everything remotely and does not hide it: "Transcription always occurs on the cloud. This is the best way for us to provide accurate, low latency transcription." All processing happens in the United States.

On training, the help centre is the document that gives you the state of play: "Privacy Mode off (standard mode): audio and transcription data may be used to evaluate, train, and improve Wispr's models. This is the default for trial and standard accounts. Enterprise and HIPAA BAA customers run with Privacy Mode on by default." The setting is presented during onboarding, pre-selected, as "Improve the model for everyone", which is also its new name in the app. If you set Flow up quickly on a Tuesday and never went back, that is the state you are in.

Turning it off takes about a minute: Settings, Data and Privacy, and switch off "Improve the model for everyone". If you want nothing stored at all, that is two settings rather than one, and the help centre says so precisely: "Zero Data Retention (ZDR) is shorthand for Privacy Mode on plus Cloud Sync off: no training and no server-side storage of dictation data."

One commitment here is better than most of the field and applies whatever you choose: "Wispr always maintains zero data retention agreements with all third-party AI providers", and those providers "do not use your data to train their models, and all shared data is deleted after 30 days". That covers the subcontractors whichever way your own switch is set.

The legal documents, both updated 19 August 2026, describe the choice rather than the default: "If you choose to share your content with us for model training, we may also use your Customer Content to train our AI models." The default state appears only in the help centre. We would rather it were in the policy, which is the same thing we said about six other products on this page.

Otter and Rev are covered above. Both train by default. Rev lets you opt out by email. Otter, on an ordinary account, does not let you opt out at all.

Descript trains only on shared data, and which state you start in depends on which Descript page you believe.

The built-ins, which most people never chose

These matter more than the paid apps, because you are probably using one right now without having picked it.

Google and Gboard

Fully on-device voice typing is a Pixel feature, not an Android feature. Google's help page is specific: "The text you speak stays on your device and isn't sent to Google servers except when you use the 'Fix it' or detailed edits features", and to get it "you must have Pixel 6 or up". On other Android phones, Google's Gboard pages do not state where the recognition happens, which is itself an answer of sorts.

Two learning mechanisms run, and their defaults point in opposite directions. Federated learning is on: "Important: Federated learning is turned on by default." It sends what the model learned rather than what you said. Audio donation is off until you agree: snippets are "sent to and stored by Google to improve speech recognition for everyone", capped at 15 seconds, kept no longer than 18 months, and "human reviewers may listen to or transcribe some snippets".

Saving voice recordings to your Google account is a third, separate thing, and it starts off: "This voice and audio activity setting is off unless you choose to turn it on." Note the trap in the same help page: "If you turn this voice and audio activity setting off, previously saved audio is not deleted." Turning it off does not clean up behind you.

Here is the mismatch. Google's privacy policy, effective 2 April 2026, says: "Depending on your settings, we can save audio recordings of voice interactions with services like Google Search, Assistant, Maps, and Gboard to develop and improve Google audio technologies." Gboard is in that list. The help page about the setting itself names only Search, Assistant and Maps.

To change things: on the keyboard, Settings, Privacy, then Voice, then Audio donations. In your account, Data and privacy, Web and App Activity, then "Include voice and audio activity".

Apple

Apple's user guide says dictation "requests are processed on your device in many languages", with "no internet connection is required". The legal notice, dated 11 February 2026, is more careful and more useful: "your device will indicate in Keyboard Settings if your audio and transcripts are processed on your device and not sent to Apple servers. Otherwise, the things you dictate are sent to and processed on the server."

The part almost everyone gets wrong is what "not stored" covers. Audio is not kept unless you opt in. Transcripts are, by default, and they are used for training. Apple's own words: "Apple may retain and use this data for up to two years to develop and improve Siri, Dictation, Search... transcripts may be used to fine-tune Siri, Search, Voice Control, Translate, and automatic speech recognition models." A subset gets human review and "may be kept beyond two years".

Opt in to "Improve Siri and Dictation" and audio joins the pile, reviewed by staff rather than contractors: "audio review being conducted only by Apple employees."

To change things: Settings, Privacy and Security, Analytics and Improvements for the opt-in. Settings, Siri, Siri and Dictation History to delete what is held. Settings, General, Keyboard to switch dictation off entirely.

Worth knowing when you weigh Apple's promises: the claims that Siri recorded people without consent produced a class action, Lopez v. Apple, filed on 7 August 2019 in the Northern District of California. The court granted preliminary approval of a settlement on 10 February 2025, the case was terminated on 14 October 2025, and class counsel put the final figure at $95 million.

Microsoft and Win+H

Windows has two speech systems and they are not interchangeable. Voice typing, the one you get with Win+H, is always remote: "Voice typing uses online speech recognition, which is powered by Azure Speech services", and it needs an internet connection. Voice access, built as an accessibility feature, runs "without an internet connection" after the first language download.

Microsoft's phrasing is where care is needed: "Voice data is sent to Microsoft only to provide the service and create text transcriptions. Microsoft does not store, sample, or listen to voice recordings without your permission." The second sentence is about storing and listening. The first says the audio goes anyway. If you press Win+H, your voice reaches Azure every time, permission or not.

Training is opt-in and Microsoft says so twice: "You can use voice typing without contributing voice clips. Contributing voice clips is optional." If you do contribute, humans are in the loop: "Microsoft employees and vendors working on behalf of Microsoft will be able to review your voice clips." The company also states it "stopped logging any voice data for product improvements beginning on October 30, 2020".

To change things: Start, Settings, Privacy and security, Speech, and switch off Online speech recognition. Be ready for the consequence Microsoft states directly, that only device-based features remain, which means Win+H stops working and Voice access becomes your dictation. Voice clips have their own control inside the Win+H toolbar under Settings.

The privacy statement, last updated July 2026, describes voice data being used "to develop and improve speech recognition accuracy" without the "with your permission" qualifier that appears in the help pages.

OpenAI and ChatGPT dictation

There is no local option. Press the microphone in the message box and "the recorded audio is sent to our models to be transcribed".

Three neighbouring features keep audio for three different lengths of time, which is why people come away confused. Dictation audio is kept "for as long as the chat is part of your chat history", and deleted within 30 days of you deleting the chat. Voice mode clips are kept 30 days. The old promise that clips are deleted straight after transcription is still there, now scoped to one mode: "If you are using our legacy standard voice mode, audio clips from ChatGPT are transcribed before we generate a response. We delete audio clips once transcription is complete."

On training, audio is off and text is on. "By default, we won't train our models with audio or video clips" from voice chats. Meanwhile, on the page about model improvement: "ChatGPT, for instance, improves by further training on the conversations people have with it, unless you opt out." Turning off audio sharing alone leaves the transcripts in play, and OpenAI says so: "To opt out of training our models entirely, please disable Improve the model for everyone."

To change things: Settings, Data Controls, then "Improve the model for everyone" and the "Include your audio recordings" toggle inside it.

The privacy policy, updated 18 May 2026, mentions audio once, in a list of content types. Every retention period and every training rule above comes from help articles that are stamped "updated 6 days ago" rather than with a date, and which are rewritten without version history. Our requests to OpenAI's domains are blocked, so we read all of these in archived copies captured on 4 June, 29 July, 19 August and 25 August 2026, and we would rather say so than imply a live reading.

What a court is currently examining

One case is worth knowing about because it tests exactly the question this page asks.

Brewer v. Otter.ai, case number 5:25-cv-06911-EKL, was filed in the Northern District of California on 15 August 2025 by Justin Brewer, and four related suits were consolidated before Judge Eumi K. Lee on 22 October 2025. The complaint alleges that Otter's meeting assistant recorded participants without the consent of everyone in the call, and that Otter "has used captured data to train its speech recognition models". It cites the federal wiretap statute, the Computer Fraud and Abuse Act and the California Invasion of Privacy Act, and seeks class certification and damages above $5 million.

These are allegations. Nothing has been proven, Otter disputes the claims, and a filed complaint is a set of assertions rather than a finding. We include it because it is the only place where statements about training on recordings are being tested by somebody other than a blog, and because the outcome will matter for every product in the table.

Picking by your constraint

You are under an NDA, or handling client matter, medical or legal material. Local only. VoiceInk, superwhisper on a local model, Handy, or Whisper through Buzz. Nothing in the cloud half of this table solves the problem, because the problem is not whether the vendor is trustworthy, it is whether the file left the building at all.

Your employer or client requires EU processing. Transkriptor states EU-only processing. Check anything else against its own documents rather than a badge, and note that at least one vendor here has two pages that disagree about geography.

You are doing research with an ethics committee. The vendor's policy is a document you will have to name in your protocol, so pick a product whose answer sits in a dated document rather than on a marketing page. That criterion alone shortens the list a lot. A researcher put the same point on Hacker News in July 2026, explaining why he built his own tool: there were "no real options for interview transcriptions with speaker detection that can be used in an university study evaluation context where you do not upload the data into a someone's else's cloud with questionable privacy policies".

You are dictating ordinary work text and just do not want to be a training set. Every cloud product here can be configured to stop. Set aside five minutes and go through them: Wispr Flow under Data and Privacy, ChatGPT under Data Controls, Windows under Privacy and security, Apple under Analytics and Improvements, Gboard under keyboard Privacy. That is the whole job.

You have a pile of recordings to turn into text and no confidentiality problem. Take the free tier of a transcription service and run your own worst audio through it before paying anyone. If you also want the no-training question settled in writing, Transkriptor is the one in this table that states it without conditions and lets you set your own deletion window, from one minute to a year. Read the paragraph above about where that statement lives, decide whether that is enough for your situation, and if it is not, go local instead. Our guide to transcribing interviews and lectures prices every route including the free ones.

How we checked this

On 27 August 2026 we read, for each product, the privacy policy, the terms of service where a separate one exists, the security or trust page, and the relevant help centre articles. We quoted them rather than paraphrasing, and we recorded the effective date printed on each document. Where a vendor's site blocked us, we used archived copies and said which snapshot and what date. Where a claim exists only in someone else's review, we left it out.

What we did not do, and will not imply: we did not intercept network traffic, decompile any app, or create paid accounts to observe default settings from the inside. This page tells you what fourteen companies have written down and where. If a company does something other than what it wrote, this method will not catch it, and no amount of policy reading would.

Four things we could not settle. Otter's enterprise help pages return an error for us, so the widely repeated claim that enterprise workspaces are excluded from training by default is unverified here; the only Otter statement we could read on it is the enterprise page telling you to contact an account manager. Descript's default state is contradicted between its own two pages. Microsoft never publishes the default value of the Online speech recognition switch. And Reddit is unreachable from where we work, so the practitioner comments on this page come from Hacker News, which skews technical.

When this page changes

Prices go stale slowly. Policies do not. Two of the documents quoted here were updated in the last fortnight, one was rewritten twice this year, and one is a class action that will produce rulings.

So this page is on a fixed re-check schedule rather than a "we'll get to it" basis. Every document is re-read on that cycle, and out of cycle whenever a vendor announces a change or a policy date moves. When a row changes, the row changes and the date at the top of the page moves with it. If you find something here that no longer matches what a vendor publishes today, tell us and we will re-read it.

Every policy, contract, help page and court record quoted above was read at its source, and each one is re-read on a fixed schedule because this is the fastest-moving material on this site. The table above is rechecked on a schedule, and the date at the top of the page is the date of the last pass. Where a vendor blocked our requests we said so and named the archived copy we used instead.