Why does talk to text censor words?

When talk to text shows asterisks where you said a swear word, the engine recognized the word, and a separate step replaced it. That step is a profanity filter. It is a product setting, so it is not a sign that the microphone or the engine failed. Whether it is on by default, and whether you can change it, depends on the product. This answer shows how the filter works, how the common products differ, and how to turn it off. The wider set of dictation errors is in the dictation mistakes guide.

The filter acts on the recognized text

A profanity setting changes the result, not the audio. Azure AI Speech, a cloud speech service for developers, has three values for its profanity option: masked, removed and raw. The raw value "includes profanity in the result" ¹. The other two values change what is returned. That is why a filter can be turned on or off without changing how you speak.

Products differ in their default and their choices

The filter is not one standard feature, and the same word can be handled three different ways.

The table shows three designs. Azure masks by default, Google masks only when asked, and Windows lets you choose in its voice typing settings. A setting with the same name can behave differently in another product, so check the one you use.

How to turn the filter off

1. On Windows, open voice typing, select the gear icon to open its settings, and change Filter profanity to allow profanity, which types it out as spoken ⁵.
2. In an Azure request, set the profanity parameter to raw ¹. This applies only to developers who call the API.
3. In a Google request, leave the profanity filter off, which is the default ³. Set it to true only when you want masking.
4. In an app with no setting, you cannot turn the filter off there. Type the word yourself, or check the output and fix the replaced words by hand before you use it.

Numbers can change too, for a different reason

Filters are not the only deliberate change. Azure's display text applies inverse text normalization, which converts spoken text to shorter forms, such as 200 for "two hundred" ⁶. You may want the words in a legal document and the digits in a spreadsheet note. Check numbers in the output, not just the words, and read more about this in the inverse text normalization glossary entry.

It depends on the purpose of the text

The right setting depends on what the text is for.

• For an exact transcript, such as a record that must match the speech, turn the filter off where the product allows it, and proofread the swear words and the numbers.
• For a shared document, the masked version may be what you want. Keep the default and check the result for words that were replaced.
• For a product with no setting, a filter changes what you see but not what was said. If the exact words matter, record the audio as well, or use a tool that offers the choice.

If the word that came out is wrong but was not a swear word, the cause is more likely recognition than a filter. For that case, see how to improve dictation accuracy. For the names and jargon that the engine does not know, see how to make dictation spell names correctly. For the checks to run on the output, see how to proofread dictated text. The terms are defined in the dictation and voice typing glossary entries, and the set of questions is part of the dictation topic.