Potok transcribes Russian speech; this site has been translated into English for convenience.
WER (Word Error Rate) is a speech recognition accuracy metric: the share of wrong words compared with a reference text. It is computed as (S+D+I)/N, where S, D and I are substituted, deleted and inserted words, and N is the total number of words in the reference. WER 10% means about 10 errors per 100 words, and Potok (potok.now) also has its own result on our GolosTest benchmark.
Developers of speech recognition programs publish WER as the main quality number: 3%, 5%, 10%. On its own it says nothing unless you know what stands behind the letters and how the figure is obtained.
Where WER comes up when you choose a dictation program
When you compare speech recognition programs, you often run into a line like “WER 5%” or “WER 10%” in a product description or an independent benchmark. It is the most common accuracy figure that developers give, but not everyone understands what it means.
WER is computed with the formula WER = (S+D+I)/N. The numerator is the sum of errors of three kinds: S (substitutions) are words recognized wrongly instead of the right ones; D (deletions) are words that recognition skipped; I (insertions) are extra words that were not in the actual speech. The denominator N is the total number of words in the reference text against which the result of speech recognition is compared.
WER 10% means about 10 errors per 100 words of the reference: some of them may be one word replaced with another, some a skipped word, some an extra insertion. The lower the percentage, the fewer differences from the reference text.
What WER (Word Error Rate) is in simple terms
Put simply, WER is a way to compare what speech recognition produced with what was actually said and to express the difference as one number. The source defines WER as a general performance indicator for speech recognition and machine translation systems: the metric applies not only to voice input but also to translation systems. How WER is applied specifically to speech recognition is covered on a separate page.
An important detail: WER is computed on a specific test set of audio recordings with a known correct text. The number applies to that set, not to the program in general; on another recording or in another language the same system will show a different figure. The metric itself does not cover automatic punctuation, which is assessed separately.
| Term | Meaning | Example |
|---|---|---|
| WER (Word Error Rate) | a speech recognition accuracy metric: the share of wrong words compared with a reference text | WER 10% means about 10 errors per 100 words |
| Speech segmentation | splitting a continuous audio stream into separate meaningful or pause-delimited fragments | segmentation helps the system recognize long speech in parts |
| Capitalization model | a separate model responsible for placing capital letters in the text | the capitalization model and the punctuation model often work as a single post-processor |
| Custom dictionary | a list of terms, names and abbreviations that a user adds by hand so that the system recognizes them more accurately | a personal dictionary helps to correctly recognize rare surnames and professional terms |
What WER does not show
WER is a metric for any speech recognition or machine translation system, not a figure for one product: it can measure both the built-in tools of an operating system and a cloud service, if there is one and the same reference set. Potok has its own result: WER 2.1% on our GolosTest benchmark. Potok is a cloud service, recognition runs on servers in Russia, and it does not work without the internet.
Frequently asked questions
WER 10% means that on average about 10 errors fall on every 100 words of the reference text: substituted, deleted or inserted extra words compared with what was actually said.
Download Potok
Works with a single key in any program. 14 days of full access, then Free with 2500 words per week or Pro at 690 ₽/mo (499 ₽/mo paid yearly). Potok is a cloud service: speech is recognized on servers in Russia, and it does not work without the internet.
Download PotokExisting recordings are transcribed by Potok's Telegram bot @potok_now_bot: send a file or a YouTube, VK Video or RuTube link and the text arrives in the chat. The Potok subscription removes the limit in the bot too.
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