How voice typing works

How speech recognition learns, and why context helps with names and jargon

Training vs everyday use, why models struggle with unfamiliar names and accents, how context hints and dictionaries steer recognition, and why dictating doesn't mean your voice trains the model.

MacWindowsiPhoneAndroid

In short

A speech model learns from huge amounts of recorded speech paired with text, then applies what it learned to your voice. It is best at what it saw often and weakest at rare names and jargon. A context hint, like the words in your dictionary, tilts recognition toward your terms without changing the model itself, and using a hint doesn't mean your recordings become training data.

Slack

You said

can you push the hotfix to kubernetes before the demo with acme

SayVelio typed

Can you push the hotfix to Kubernetes before the demo with Acme?

Training and use are two different moments

Machine learning lets a model find patterns from examples instead of following hand-written rules. For speech, those examples are recordings paired with what was said. Training adjusts the model's internal parameters across a huge number of such examples. Inference is what happens when you dictate: the finished model is applied to your new recording.

The two are separate. Dictating a sentence doesn't by itself teach the model anything. Whether a provider ever uses customer audio for training is a policy question, not a technical necessity. SayVelio doesn't train AI on your data; see privacy in voice typing.

What a model is good at, and where it struggles

A model reflects what it was trained on. A common word said clearly in a quiet room is easy. Things get harder when your audio differs from the training data:

  • Rare words: surnames, product names, internal project names, new slang.
  • Accents and speaking styles that were less common in the training data.
  • Conditions: echo, distant microphones, several people talking.
  • Mixed languages within one sentence, which SayVelio handles by detecting the language on every dictation (see dictating in two languages).

That's why quality has to be measured on recordings the model never saw during training. A model that looks perfect on its own training data can still stumble on your voice.

More data isn't automatically better

It's easy to assume that more features, more measurements or more parameters always mean better results. In practice, irrelevant or redundant information can make a model learn noise instead of patterns that generalize. Researchers test models on held-out recordings precisely to check whether an improvement survives on new speech. For you, only one number matters: how well it transcribes your next sentence.

Two ways to fix a stubborn word

Context hintFine-tuning
What changesWhich candidate words are favored for this requestThe model's parameters, through extra training
How fastInstantly, per dictationTakes training data and time
Who does itYou, by adding wordsThe model provider
GuaranteeNone, but very effective for namesNone either

A context hint is like telling a person before a meeting: "you'll hear the names Kubernetes and Acme". It doesn't change their hearing, it changes their expectations.

How SayVelio uses your dictionary

Your SayVelio dictionary works as that kind of hint at two points:

  1. During recognition, your words are passed to the speech model as context, so it's more likely to choose them when the sound fits.
  2. During cleanup, if it's on, the dictionary helps fix spellings that recognition still got wrong, including the variants you listed for how a word sounds.
Dictionary words act as hints during recognition and during cleanup
Your dictionary steers both recognition and cleanup toward your words

Neither step is a guarantee, and a short, focused dictionary works better than a long one, because hundreds of hints dilute each other. Add the words that actually come out wrong, and let unused ones drop off.

Cleanup is a second model, with its own risks

Recognition writes down words; an optional cleanup step rewrites them into readable text. Cleanup uses a language model, which is great at punctuation and removing "um"s, but a model that rewrites too freely can also change meaning. SayVelio's cleanup is instructed to keep your words and your language, never translate and never answer what you said. Still, for anything important, compare the result with what you meant.

Questions

Does SayVelio learn my voice over time?

No. The recognition model doesn't adapt to your voice, and your recordings are not used for training. What adapts to you is your dictionary, your cleanup prompt and your app styles.

Why does adding a word not always fix it?

A hint makes a word more likely, it doesn't force it. If the word still comes out wrong, add how it sounds as a variant, or check that the word isn't drowned out by a very long dictionary.