Speech to text in Russian: how it works and which programs there are

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

Speech to text in Russian is recognizing Russian speech into text, and it works differently from program to program: some models are trained on many languages at once, others specifically for one language. Potok (potok.now) uses its own model for Russian speech, inserts text straight into the active field without a separate export, and on the free plan gives 2500 words per week.

Speech to text in Russian appears in two different contexts: as the general name of the speech recognition technology and as a search query of a person choosing a specific program. Behind the phrase there is usually one of three things: dictation right into an input field, transcription of an already finished recording, or voice control by commands. Program websites word this vaguely, and from the first click it is not always clear what is in front of you. The difference between the approaches determines which speech-to-text recognition program fits the task at all, and it is worth starting with that, not with comparing accuracy figures on websites.

How speech recognition in Russian works

The mechanics are the same in almost all of them: a microphone records sound, a neural network turns it into text, and the program inserts the result where the cursor is or saves it as a separate file. The difference is where this neural network computes and for which language it is trained first of all.

Some programs work in the cloud: the sound goes to a server, is recognized there, and the text comes back. SpeakFlow and Diktuy are cloud services with payment in rubles directly, although the country of their data center is not explicitly disclosed on their websites. Diktuy states an accuracy of 95–98% on clean Russian speech; this is the service's own figure and cannot be checked independently, but its source is specific. Talkpad and Lispr are also cloud-based, only built differently: Talkpad is a multilingual automatic language detector with over a hundred languages and no separate model for Russian, while Lispr recognizes through the company Groq on servers in the USA.

Another part works locally: the sound does not leave the computer at all. Voisty uses models such as Whisper Large V3 Turbo or NVIDIA Parakeet, WhispeRu recognizes offline with no network access, and GigaType is built on GigaAM, a Sber neural network trained specifically for Russian. Molvi is also local, with support for Russian and English and automatic detection of which of them is being spoken.

Potok takes a third approach: the cloud, but the model is its own and trained specifically for Russian speech, not for dozens of languages at once, aimed at names, terms and Russian-English phrases like “otprav draft v CRM” (send the draft to CRM). Response is about 0.4 seconds, and the personal dictionary and dictation history are tied to the account, not to one computer. A multilingual model kept in the cloud to support many languages at once is an engineering compromise, not a quality indicator in itself: it may handle Russian superbly or mediocrely, and without an honest test on real phrases you cannot know.

Where dictation saves time and where it gets in the way

By voice you can type faster than by hand: about 900 characters per minute against about 200 when typing. The difference is especially noticeable on long letters, office memos and drafts that are edited afterwards anyway. If the typical task is to speak a paragraph and fix a couple of words, dictation almost always wins on time, not only on the feeling of convenience.

Where the text is short and exact (a one-word reply in a chat, a phone number, an email address), the gain is less noticeable: while you open the microphone and wait for recognition, you could have typed the same by hand. Voice wins on volume, not on reaction speed for a single phrase.

A separate story is a noisy place. An open-plan office, a cafe, a car with a running engine: background noise hinders recognition in any program, and what helps here is not the choice of a specific service but a microphone physically closer to the lips, for example one in headphones, not the one built into the laptop.

Another scenario is switching between several windows at a time: a letter in one app, a chat in another, a task in a task tracker. Dictation inserts text where the cursor is, in any of them, without a separate step of copying from its own window. This saves seconds on each switch, and over a working day of dozens of such switches noticeable time adds up, even if each one seems a trifle on its own.

Work with professional terms and client names is where the difference between programs is especially tangible in practice. A personal dictionary that remembers such words saves time on corrections: without it you have to either spell the term out each time or clean the text by hand afterwards of wrong recognitions. Not all the programs mentioned above have such a feature at all: ask the specific service, and do not count it as standard.

When it is better to type or take another tool

If you need to transcribe an already existing recording (a meeting, an interview, a voice message from a messenger), speech to text in the sense of dictation does not fit at all: Potok recognizes only what you say right now, not a file from a disk; for ready recordings Potok has the Telegram bot @potok_now_bot. For transcribing ready recordings you need a separate tool that takes an audio file as input.

Potok is a cloud service, recognition runs on servers in Russia, and it does not work without the internet. For those who need to work on a plane, in a room with no network or at a site with a ban on cloud processing, a cloud option will not fit in principle: only a program with recognition right on the device will do, for example one of the local solutions covered above.

Potok has no open API for developers: it is a ready app for the end user, not a construction kit for embedding in your own software. There are no mobile versions for Android and iPhone either: clients only for Windows and macOS, and part of the “speech to text Russian” queries are actually searched for a phone: voice notes on the go, where there is no computer nearby. A more basic question, whether a computer has voice input at all without installing anything, is covered separately.

Frequently asked questions

Related software: terms compared

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