Autoreplace, autocorrect and voice input: how they get in each other's way and help each other

Potok transcribes Russian speech; this site has been translated into English for convenience.

Autoreplace and autocorrect react to text typed letter by letter, while voice input inserts an already recognized phrase whole, so some of the system's correction rules simply do not fire on it. Sometimes this helps: there are fewer typos from the hand. Sometimes it gets in the way: the system may still swap a term or a name for its own variant. Potok (potok.now) reduces this with a personal dictionary of terms, and its response time is about 0.4 seconds.

Autoreplace puts a preset variant in place of a typed abbreviation, and autocorrect fixes typos, case and punctuation on the fly. Voice input works differently: it does not type letters one after another, but inserts a recognized piece of text at once into the field where the cursor is. Because of this difference the three mechanisms do not always act in agreement: in some places they help each other, and in others they create confusion that is easier to foresee in advance than to untangle later.

How voice input works

Most voice input programs follow one scheme: pressing a hotkey turns on microphone recording, the speech goes to recognition, in the cloud or locally on the device itself, and the finished text is inserted into the field where the cursor is. In Potok the default hotkey is Ctrl+Space: hold, speak, release, and the text appears in the active app.

An important detail for the topic of autoreplace is exactly how the text gets onto the screen. In most programs autoreplace and autocorrect watch letters as they are typed: a rule that fires replaces the typed combination of characters right after the user finishes typing it. Voice input bypasses this process: the recognized phrase arrives in the field as one block, not character by character, so some rules designed for character-by-character typing simply do not get a chance to fire. Other rules still fire, though: those that check the already finished text as a whole rather than the moment a letter is typed.

Voice input programs also differ in recognition architecture, and this too affects how fast text appears in the field. SpeakFlow, Talkpad and Lispr recognize speech in the cloud: the sound goes to a server and comes back as text. GigaType, Molvi, Voisty and WhispeRu, on the contrary, recognize speech right on the device, without sending sound outside. This choice determines whether the internet is needed for dictation, but the principle of inserting the finished text into the field is the same.

Where dictation saves time and where it gets in the way

Voice input saves time where autoreplace is usually useless: in long sentences with a free vocabulary, where there are no preset abbreviations to substitute. There are also fewer typos from the hand in dictated text, because letters are not typed by hand, and so there is less ground for autocorrect to fire on typos.

Voice input gets in the way where autocorrect checks the already finished text as a whole rather than the moment of typing. A specialized term, a rare name or an abbreviation that is not in the program's dictionary can be recognized correctly and then replaced by autocorrect with a similar dictionary word. From the app's point of view this looks like an ordinary typo fix, but in fact it spoils correctly recognized text. The same applies to autoreplace: if an abbreviation that normally expands into a phrase when typing happens to match a dictated word, it may expand where it is not needed.

A personal dictionary of terms and names, as in Potok, reduces this conflict: words from the dictionary are recognized in their original form and fit less under the autoreplace rules designed for a general vocabulary. How models recognize terms in Russian speech at all is in the material on voice input for Russian speech.

When it is better to type or take another tool

Manual typing with autoreplace turned on is more convenient where the text consists of short, often repeated phrases with preset abbreviations: addresses, template replies, legal wording. Voice input gives no gain for such tasks: transcribing and setting up exceptions for autoreplace takes longer than simply typing the familiar abbreviation by hand.

It is also worth remembering the limits of Potok itself: it is a cloud service, speech recognition runs on servers in Russia, and it does not work without an internet connection. It has no voice commands for editing text that is already inserted: if autoreplace swapped something, you have to fix the text by hand, as with any other dictation program. For texts where absolute accuracy of specific terms is needed with no right to substitution, it is worth checking the result as a whole right away rather than trusting dictation and autoreplace at the same time.

Frequently asked questions

Related software: terms compared

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