Potok transcribes Russian speech; this site has been translated into English for convenience.
A dialect is a regional variety of a language with its own features of pronunciation and vocabulary. It matters in voice input because speech recognition models are trained mostly on the literary norm, not on dialect speech. A model trained on the literary norm may cope worse with a strong dialect, and recognition accuracy drops in that case.
"Dialect" in a conversation about speech recognition is not the same as "accent", although the terms are often confused. Let us look at what a dialect means exactly, what the literary norm of a language has to do with it, and why it affects the accuracy of voice input.
What a dialect is: the definition
A dialect, by the definition in English Wikipedia, is a variety of a language spoken by a particular group of people; a regional dialect is a language variety characteristic of a specific geographic area. It is a variety within one language, not a separate language: a dialect has its own features of pronunciation and vocabulary, that is, not only an accent but also some individual words and phrases characteristic of exactly this language group.
For voice input this is not an abstract linguistic question: speech recognition models are trained on a body of recordings, and in this body the literary norm of the language usually dominates over dialect speech. The term's category in the glossary is speech-tech: it concerns how exactly a system turns sound into text, not the design of a program or its interface. The more a dialect differs from this literary norm in unfamiliar words and pronunciation, the more noticeable it can be in recognition accuracy. This is how a dialect differs from an accent: an accent is mostly features of pronunciation, while a dialect is also part of the vocabulary, which may be unfamiliar to the system.
| Term | Meaning | Example |
|---|---|---|
| Dialect | a regional variety of a language with features of pronunciation and vocabulary that can affect recognition accuracy | a model trained on the literary norm may cope worse with a strong dialect |
| Accent and recognition | the effect of a speaker's pronunciation features on speech recognition accuracy | a strong regional accent can increase the number of ASR errors |
| Number normalization | conversion of spoken numerals into digit notation during speech recognition; in the industry this post-processing step is called inverse text normalization (ITN), while “text normalization” (TN) is the opposite direction, turning digits into words for speech synthesis | “dvadtsat pyatoye sentyabrya” (twenty-fifth of September) becomes “25 September” |
| Wake word | a special word that the device recognizes constantly in the background to turn on a voice assistant | “OK, Google” is a typical example of a wake word |
An example from dictation
A model trained on the literary norm may cope worse with a strong dialect. In practice this means that dictation with pronounced dialect words and pronunciation may be recognized less accurately than speech close to the literary norm.
Similar logic applies to other terms from the same area: the WER (Word Error Rate) metric serves exactly to measure the final recognition accuracy in such cases. More about the technology of turning speech into text is on the page about ASR (automatic speech recognition), and the full list of concepts is collected in the glossary of speech recognition terms.
Where Potok has limits here
Potok's recognition is in the cloud: the service runs on servers in Russia and does not work without the internet.
Frequently asked questions
Not quite: by definition, a dialect is a variety of a language with features of pronunciation and vocabulary, that is, of individual words, and not only a manner of speaking. Both matter for speech recognition: unfamiliar words lower accuracy just as unfamiliar pronunciation does.
If you want to compare for yourself
Take one of your own tasks and dictate it in Potok. The first 14 days have no limits and need no card. Potok is a cloud service: speech is recognized on servers in Russia, and it does not work without the internet.
Download PotokOne subscription, two tools: dictation in any program on your computer and @potok_now_bot in Telegram, which transcribes voice messages, audio and video.
Built-in dictation and Potok
| Tool | Where it works | Russian language | Where speech is recognized |
|---|---|---|---|
| Windows voice typing (Win+H) | Windows 10 and 11 | Windows 11 only | Microsoft cloud, does not work offline |
| macOS Dictation | macOS | yes | on the device or on Apple servers |
| Potok | Windows, macOS, Linux | primary language | cloud, servers in Russia, does not work offline |
Primary sources: en.wikipedia.org · support.microsoft.com