Voice input and personal data processing: what to check before choosing a service

Potok transcribes Russian speech; this site has been translated into English for convenience.

Before choosing a voice input service, check where recognition happens: on a server or right on the device. If on a server, the country of the data center and compliance with the 152-FZ law for the data of Russians matter. Potok (potok.now) uses cloud recognition with servers in Russia, and for business, servers inside the company perimeter are available.

Dictation goes through a program that hears everything that is said: work correspondence, phone numbers, sometimes passwords, if the field for them is not protected. Where the voice recording and the recognized text go after that is not an idle question but a quite specific one: with some programs the audio flies off to a server, with others it never leaves the computer. The difference affects both privacy and whether you need the internet to work.

What happens to the voice in cloud services

Cloud recognition is simple: the voice recording goes to the company's server, is turned into text there and comes back to the app. For privacy two things matter here: which country receives this data and what the service says about itself at all. Potok uses cloud recognition, with servers in Russia, and data handling complies with the 152-FZ law. For business there is a separate option with servers inside the company perimeter, single sign-on through SSO and a shared dictionary on request.

Not all cloud services name the country of their servers. SpeakFlow and Diktuy, according to their sites, recognize speech in the cloud but do not disclose which country the data center is in. Talkpad is in a similar situation: the legal entity itself is registered in Australia, and the exact location of the recognition servers is not stated in the privacy policy, according to the site. Lispr discloses this directly: recognition there is done by a third-party service, Groq in the USA, which means the voice and text pass through infrastructure outside Russia.

What recognition right on the device gives

Local recognition works on the opposite principle: the audio does not leave the computer at all, and all processing runs on its processor or graphics card. There is no need to find out which country the server is in, because there is simply no recognition server, and so no extra link to which you have to trust your voice recording. The other side is the load on the computer itself: recognition quality and speed depend on its power.

Voisty recognizes on the device offline, with local models such as Whisper Large V3 Turbo and NVIDIA Parakeet. WhispeRu also works offline, on Windows only. GigaType is built on the GigaAM neural network from Sber, trained for Russian, and also recognizes right on the device. Molvi has a similar scheme: the best quality comes on an NVIDIA graphics card, and without it the program switches to a faster model for the processor. None of the four programs sends audio to a third-party server.

How to decide in five minutes

Check a few items in a row, without going deep into legal texts (the full list of programs to compare is in the selection of voice input programs):

• open the service's privacy page or FAQ and find a direct answer to where recognition happens, on a server or on the device;
• if the service is cloud-based, look for the name of the country of the servers; if it is not on the site, that is also information, and you will have to either trust the service or take a local option;
• for work correspondence, find out whether there is a separate plan with servers inside the company or SSO, which for Potok is the business option on request;
• do not dictate passwords and other protected data aloud: this applies to any voice input program, not just one particular service.

Frequently asked questions

Who Potok will not suit

Potok is a cloud service, speech recognition happens on servers in Russia, and it does not work without an internet connection: it has no local mode in which audio stays on the computer. If a company forbids sending voice and text to external cloud services altogether, the regular plan will not fit. You need a separate business option with servers inside the company perimeter, and its terms are set on request, not arranged as an ordinary subscription. For those to whom it is essential that a voice recording physically never leaves the device, it is worth looking straight at local recognition programs rather than cloud ones.

Related software: terms compared

Built-in dictation and Potok

Primary sources: speakflow.ru · speakflow.ru · speakflow.ru · blog.speakflow.ru