I use software for speech-recognition to prepare documents and e-mail. I also use it to control some functions of the computer. I do this to avoid excessive use of the keyboard and mouse, which for me can cause discomfort and pain.
The software I use for speech recognition, either ViaVoice or Dragon Naturally Speaking, is not perfect.
Although it is an enormous help in quickly producing documents without excessive use of the keyboard, it does make a few mistakes.
Generally the software goes by the sound of words, augmented by some frequency tables which tell it how likely a given word is, given that some other nearby words have already been recognised. Apart from that, the software doesn't actually understand the language, and therein lies the difficulty.
The most common errors are either non-speech utterances (e.g. breath noises) which were interpreted as small words, or the substitution of one word for another similar-sounding (a homophone). Often a nuisance, these errors can also be amusing. Susan Fulton's voice recognition site has a collection of these.
Even experienced speech software users always proof-read their own text, but the best proof reader is invariably another person.
Sometimes these speech-recognition errors, or speakos, won’t have been spotted before I send the text on.
If the text I sent you looked strange, try reading it out aloud, or better yet get somebody else to read it to you.