Restoring punctuation and capital letters: how voice input places marks without the word "zapyataya" (comma)

Potok turns speech into text in 33 languages, including English and Russian.

Restoring punctuation and case means the program itself places periods, commas and capital letters by pauses and the sense of the phrase, not by separate command words like "zapyataya" (comma) or "tochka" (period). Most modern dictation services work this way. Potok (potok.now) is one voice input option, with a free limit of 2500 words per week on one device.

Voice typing used to require saying punctuation marks aloud: say "zapyataya" (comma) and a comma appeared, forget it and the sentence stuck together into one stream of words without pauses. Modern dictation services have removed this command: the program itself decides where to put a period, a comma or a capital letter, going by the pauses in speech and the structure of the phrase, not by a separately spoken command word. Let us look at how this works and where such automation still makes mistakes.

How punctuation restoration works

There are two fundamentally different approaches. The first is command-based: the program recognizes the word "zapyataya" (comma) as an ordinary word of speech but does not insert it into the text and turns it into the symbol ",". The words "tochka" (period), "novaya stroka" (new line), "voprositelnyy znak" (question mark) and a dozen similar commands work the same way. The approach is simple and predictable, but it requires keeping a list of commands in your head and interrupting your thought to say a service word aloud instead of a meaningful one.

The second approach is automatic punctuation restoration, and it works differently. The program first gets from speech recognition a stream of words without a single punctuation mark, and then a separate model analyzes the resulting text and decides where by sense and grammar a period, a comma, a dash or a capital letter should stand, roughly as a person reading someone else's transcript without marks would do. The audio recording itself gives an extra signal: a long pause between phrases more often means the end of a sentence, and a characteristic intonation at the end means a question. The user does not need to say anything special: the word "zapyataya" (comma) in the text will stay the word "zapyataya" (comma) if it really was part of the phrase.

Most modern dictation services work in exactly the second way: automatically, without a separate command for punctuation marks. This applies first of all to continuous dictation sentence after sentence: if you speak a long text without stopping, the model marks it up as a whole, going by the structure of each fragment separately and not only by the pauses between words.

The automatic approach has a cost: it is not perfect. The model can get the place of a comma wrong in a long sentence with a complex structure, fail to recognize a question asked in a flat voice without a question intonation, or put a capital letter in the wrong place after an abbreviation. Because of this, long and conversational phrases usually need a quick check after dictation, while short and structurally simple sentences are almost always ready at once. How models recognize Russian speech at all is covered in more detail in the material about voice input for Russian speech (in Russian).

Where automatic punctuation helps and where it gets in the way

Automatic punctuation saves the most time where the text is ordinary connected speech: letters, messages in messengers, notes, drafts of documents. Here you can simply speak as you usually speak aloud, without being distracted by having to say each comma as a separate word, and it is in this scenario that the automation copes best, because the structure of sentences is predictable and the model has seen similar constructions many times.

The automation copes worse where punctuation does not follow ordinary grammar but has a special meaning: lists, where an enumeration is better set out on new lines and not with commas inside one paragraph; addresses and phone numbers, where hyphens and spaces matter in specific places; email addresses and links, where a period is part of the address and not the end of a sentence, and the model sometimes confuses one with the other. In such fragments it is easier either to check the text by eye right after dictation or to insert the complex part by hand.

A separate difficulty is question and exclamation sentences spoken in a flat voice without the characteristic intonation: the model goes, among other things, by how the voice sounds, and if a question sounds like a statement, the model may not put a question mark. The same applies to quotation marks and brackets: in spoken speech there is no clear signal where to open and close them, and not every program guesses their place automatically. Such fragments usually have to be finished by hand after dictation regardless of the specific service.

When it is better to type or check by hand

When the accuracy of punctuation is critical, in legal documents, financial reports, program code, where an extra comma or a missing period changes the meaning or breaks the syntax, you should not rely on the automation entirely: such a text has to be proofread in any case, and voice input here saves time on the typing itself and does not cancel the check.

Potok is a cloud service, with recognition on servers in Russia, and it does not work at all without the internet. This is worth considering in advance: if you need dictation where there is no network, no automatic placement of marks will help, because speech recognition simply will not start. Those for whom an offline mode matters more than punctuation quality should look at local programs like Molvi, GigaType or WhispeRu, which recognize speech right on the device. You can check which of them copes better with your particular texts only on your own phrases during a trial period.

Frequently asked questions

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