Question mark, exclamation mark and ellipsis: does voice input cope without saying them aloud

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

Voice input usually places the question mark itself, from question words and the structure of the phrase, without a spoken command for the question mark. The exclamation mark and especially the ellipsis are harder for the automation: there are no clear text cues for them, and the result depends more on intonation and the length of a pause. Potok (potok.now) with a free limit of 2500 words per week places marks just as automatically, but it is better to recheck the ellipsis and exclamation marks after dictation.

Three marks, the question mark, the exclamation mark and the ellipsis, behave differently in automatic dictation, and the reason is not the whim of a particular program but how many cues for each mark there are in the phrase itself. A question usually has a question word or a characteristic word order, while an exclamation and an ellipsis have only intonation and a pause, and that signal is often not enough to place the mark with confidence.

How the model decides which mark to place

After speech has been turned into words, a separate model analyzes the resulting text and decides where a period, a comma, a question mark or an exclamation mark is needed, without the user saying the names of the marks aloud. Each of the three marks discussed has its own set of cues, and they are not equal.

The question mark is helped by the vocabulary itself: the Russian question words “kto” (who), “chto” (what), “gde” (where), “kogda” (when), “pochemu” (why), the particle “li” and a characteristic word order occur almost exclusively in questions, and the model recognizes them confidently in the text, even before taking intonation into account. It is harder with questions without a question word, “ty pridyosh” (“you'll come”) instead of “pridyosh li ty” (“will you come”): here the only cue in writing disappears, and the decision depends entirely on whether the voice rose at the end of the phrase.

The exclamation mark in Russian is helped by almost nothing in the words themselves: “kak krasivo” (“how beautiful”) can be said with delight or flatly, and in text it looks the same. The only signal is acoustic: volume, tempo, a characteristic rise and sharpness of pronunciation. If an emotional phrase is dictated in an even voice, the model is more likely to place a period than an exclamation mark, and conversely an ordinary phrase said loudly and quickly may get an exclamation that was not meant.

The ellipsis is the hardest: it means not the end of a thought but its break or a pause before a continuation, and telling such incompleteness from an ordinary long pause between sentences is almost impossible without understanding the meaning of the whole conversation. Of the three marks, this is the case where the model most often chooses the more predictable option, a period or nothing at all, and leaves the ellipsis to the user.

Where the marks are placed correctly and where you have to edit

The automation copes best with direct questions that start with a question word: “chto novogo” (what's new), “kogda vstrecha” (when is the meeting), “pochemu ne rabotaet” (why it does not work). Here there is an unambiguous text cue, and the question mark is almost always in place without additional checking. The list of programs that use such automatic markup is collected in the overview section.

It is harder with short emotional remarks and exclamations, especially if they are dictated in a calm voice simply by meaning, “otlichno, vopros zakryt” (great, the question is closed), “vot eto da” (wow). Here the result is less predictable: the same text may get a period in one dictation and an exclamation mark in another, depending on the intonation with which it was said at the moment.

The hardest are texts where punctuation carries literary or conversational meaning: remarks with a thought broken off mid-word, irony conveyed by an ellipsis after a short phrase, deliberate build-up through several exclamation marks in a row. Voice input almost never reproduces such markup by itself; it is added by hand after the main text is recognized. The specifics of recognizing Russian speech in particular, including intonation and grammar cues, are covered in more detail in the article about voice input for Russian speech (in Russian).

When to check the text by hand

If ellipses, rows of exclamation marks or fine intonation markup matter in the text (fiction, social media posts, correspondence with an emotional tone), you should not rely on the automation: it is faster and more reliable to place such marks by hand after dictation than to redictate a phrase several times hoping for the right mark.

Potok is a cloud service, with recognition on servers in Russia; it does not work without the internet, and this should be considered separately from the question about punctuation: if you cannot dictate to check a specific phrase because there is no network, neither the placement of marks nor recognition will happen at all. Where exact markup of questions, exclamations and ellipses matters more than typing speed, proofreading the text after dictation is a mandatory step, not a spare option.

Frequently asked questions

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