Potok transcribes Russian speech; this site has been translated into English for convenience.
The signal-to-noise ratio shows how much louder the voice is than the background noise in a recording: the closer the noise is in volume to the voice, the more often the program confuses words, and this does not depend on how good the recognition model itself is. The WER of 2.1% that Potok states on the GolosTest benchmark was measured on clean speech; in a noisy room the result will be worse for any program.
A microphone records not only the voice: it also catches at the same time the hum of a fan, noise from the street, the sound of a neighbor's keyboard and dozens of other sources of sound around. The signal-to-noise ratio is how much louder the voice is than all of this background in a specific recording. The higher it is, the more confidently the program separates speech from interference; the lower it is, the more often it takes noise for part of a word or, the other way around, loses part of a word in the noise.
How noise gets in the way of recognition
Any speech recognition program receives as input not a clean voice but an already finished mixture of voice and background sound from the microphone. They can be separated after recording only partly, and the stronger the noise was at the moment the phrase was spoken, the less useful signal from the voice is left for analysis. This happens at the very first step, before the sound gets into the recognition model, whether it is local on the device, as in GigaType, or in the cloud on a server. So the problem of a low signal-to-noise ratio concerns any program, regardless of how and where it is built further on.
The boundaries between words and quiet unstressed sounds inside words are especially sensitive to noise: exactly there the difference between voice and background is minimal, and it is hardest for the model to decide where one word ends and the next begins. Loud steady noise like the hum of an air conditioner interferes less than short sharp sounds: a door slam, a dog barking, someone else’s remark nearby. They cover part of the speech exactly at the moment it is recorded, and this part of the signal cannot be restored by any review of the result afterwards.
Distance to the microphone works in a similar way: the farther the mouth is from the microphone, the quieter the voice becomes relative to the same background noise, that is, the signal-to-noise ratio falls even without a single extra source of sound in the room. So a headset with a microphone near the mouth or a separate microphone on the desk usually gives a cleaner recording than a laptop built-in microphone half a meter from the face.
Which setting helps dictation and which gets in the way
A quiet room with the door closed and the TV off is almost ideal conditions: there is little background noise, and the signal-to-noise ratio stays high throughout the dictation. Here you can count on an accurate result in almost any program, if the microphone is fine and not too far from the speaker. How exactly models adapt recognition to Russian speech is covered separately in the article on voice input for Russian speech.
An open-plan office, a cafe with background music and conversations at neighboring tables is a much harder environment: extraneous voices interfere with recognition especially, because in frequency and character of sound they are closest to the voice of the person dictating, and it is harder for models to separate them than steady noise from cars or the street outside the window.
Dictation outdoors or in transport adds variable noise as well: gusts of wind straight into the microphone, engine hum, random sharp sounds, which change from second to second and do not let the signal-to-noise ratio stay stable. In such conditions it is not so much the choice of program that helps as physically bringing the microphone closer to the mouth: a headset or wired headphones with a microphone noticeably raise the share of voice in the recording compared with a microphone on the desk or in the laptop body.
| Task | Characters | Typing (200 chars/min) | Voice (900 chars/min) |
|---|---|---|---|
| chat reply | 300 | 2 min | 0.5 min |
| 1500 | 8 min | 1.7 min | |
| post or meeting minutes | 4000 | 20 min | 4.4 min |
| article | 10000 | 50 min | 11.1 min |
When a low signal-to-noise ratio cannot be worked around
Potok is a cloud service: recognition runs on servers in Russia, and it does not work without an internet connection. It has no offline mode, so where you need dictation in a noisy place without a reliable network, for example on a construction site or on the road with poor coverage, plan both the microphone and the internet connection in advance, and do not count on the program alone.
There is noise that no microphone and no recognition model can handle: several people speak at the same time and at about the same volume, which makes the voices indistinguishable from each other by sound. In such a setting it is easier to step aside and dictate the phrase again than to rely on the program to pick out the needed voice from the general hum.
Frequently asked questions
Noise cancelling in headphones removes noise for the person who listens, but it does not always affect what the microphone of the program records. The quality and position of the microphone itself matter more: a headset with a microphone near the mouth usually gives a cleaner voice recording than a laptop built-in microphone at a distance.
It is not only about the volume of noise but also about its character: the steady hum of a fan interferes less than short sharp sounds like a door slam or someone else’s voice nearby. They get into the recording exactly at the moment a word is spoken and cover it for the microphone.
Potok states about 900 characters per minute by voice against about 200 when typing, roughly 4 times faster. This is the product's own estimate, not an independent measurement.
Built-in system tools (Win+H in Windows, Dictation in macOS, Voice Typing in Google Docs, Gboard on Android), browser voice extensions, and dedicated desktop or web services (Potok and similar ones) that work on top of any app through a hotkey.
Voice input is converting speech into text on the fly: the user speaks, and the program transcribes the sound with speech recognition technology (ASR) and inserts the resulting text into the active field: email, messenger, document. In Potok it works through a hotkey: hold it, speak, release, and the text is inserted.
Dictation pays off most for long texts written off the cuff: draft emails, notes, replies in messengers, lecture notes, and also when your hands are busy or tired (for example with RSI). For short precise edits (one-word commands, code) typing is often faster.
How to start
Potok is available for Windows, macOS and Linux. For two weeks everything is open for free, after that 2500 words per week without paying. Potok is a cloud service: speech is recognized on servers in Russia, and it does not work without the internet.
Download PotokYou dictate in the app, and recordings and voice messages are transcribed by @potok_now_bot in Telegram. 60 minutes per month are free; the Potok subscription removes the limit in the bot and in the app.
Related software: terms compared
| Service | Platforms | Free | Price | Payment from Russia | Where speech is recognized |
|---|---|---|---|---|---|
| Potok | Windows, macOS, Linux | Free: 2500 words per week; full access for the first 14 days, no card needed | Pro 690 ₽/mo, 499 ₽/mo when paid yearly | Russian bank cards, rubles | cloud, servers in Russia; internet required |
| SpeakFlow | Windows, Web | 30 minutes of dictation per month, free forever | 690 ₽/mo (Personal (unlimited)) | rubles directly | cloud (the data center country is not disclosed on the website) (not confirmed) |
| Diktuy | Windows, macOS | 30 minutes of dictation per month, free forever | 239 ₽/mo (Profi (monthly equivalent when billed yearly)); 2870 ₽/yr (Profi (year, billed in full)); 479 ₽/mo (Max (monthly equivalent when billed yearly)) | rubles directly (not confirmed) | cloud (the data center country is not disclosed on the website) (not confirmed) |
| Talkpad | macOS, Windows, Android | 2500 words/week on desktop, 1500 words/week on mobile, free forever; plus a 14-day Pro trial | $6/mo (Pro (billed yearly, $72/yr)); $8/mo (Pro (monthly billing)) | unknown (not confirmed) | cloud (the legal entity is registered in Australia; the website is hosted on Cloudflare; the exact location of the ASR servers is not disclosed in the privacy policy) (not confirmed) |
| Voisty | Windows, macOS | 5 minutes of transcription (to judge the quality) | $6.7/mo (Pro); $67/yr (Pro (year)); $99/one-time (Lifetime) | Visa/Mastercard/Amex cards, PayPal; whether Russian cards are accepted is not stated (not confirmed) | device (offline, local models such as Whisper Large V3 Turbo / NVIDIA Parakeet) |
| WhispeRu | Windows | 7-day free trial (no card) | 290 ₽/mo (monthly license); 1990 ₽/yr (yearly license); 4990 ₽/one-time (lifetime license) | rubles directly (payment via YooKassa) | device (offline) |
| GigaType | Windows, macOS | free with no limits: open source under the MIT license (Copyright (c) 2026 Disrupt Builders), all features are available without payment | — | not needed: there are no paid plans | device (offline, Sber's GigaAM neural network) |
| Molvi | Windows, macOS | free forever, no accounts and no payment | — | not needed: the product is free; voluntary donations are accepted via Telegram Tribute or by card on the website, amounts are not fixed | device (offline; best quality on an NVIDIA GPU, otherwise a fast model on the CPU) |
Primary sources: speakflow.ru · speakflow.ru · speakflow.ru · blog.speakflow.ru