How voice typing works

Preparing your voice for recognition: noise suppression, echo control and voice detection

What happens to audio before speech recognition: gain control, noise suppression, echo cancellation, voice activity detection and spectrograms, and why "cleaner" audio isn't always better.

MacWindowsiPhoneAndroid

In short

Before recognition, audio may be leveled, cleaned of echo and noise, and cut into the parts that contain speech. Each of these steps has one narrow job, none of them understands words, and none is required. Strong noise removal can even hurt, because it can erase the quiet consonants that tell words apart.

Gmail

You said

hi tom the invoice is attached let me know if the numbers look right

SayVelio typed

Hi Tom,

The invoice is attached. Let me know if the numbers look right.

Processing the signal is not recognizing speech

Signal processing works on sound, not on language. It can make speech louder, quieter or clearer, or decide where speech starts and stops, but it never decides which words you said. That job belongs to the recognition model, described in how speech recognition works.

Which steps an app uses depends on the situation. A phone call needs echo cancellation because the other person's voice comes out of your speaker. A dictation app with nothing playing may not need it at all.

Four common tools, four different jobs

ToolWhat it doesWhat it can't do
Automatic gain controlRaises or lowers the recording levelFix a voice that was never captured clearly
Noise suppressionReduces steady background sounds like fans or trafficSeparate two people talking at once
Echo cancellationRemoves your own speaker's sound returning to the microphoneHelp when nothing is playing
Voice activity detectionEstimates which moments contain speechTell what was said

Echo cancellation is the most specific of them. It knows exactly what your device is playing, because it gets a reference copy of that sound, and subtracts the part the microphone picks up again. That's different from noise suppression, which has to guess what is noise.

Voice activity detection: finding the speech

A voice activity detector (VAD) looks at short slices of audio, often 10 to 30 milliseconds, and estimates how likely each one contains speech. The simplest version just checks loudness, but a loud door slam crosses a loudness threshold and a quiet speaker can fall below it. Better detectors estimate a speech probability per slice, merge nearby speech regions, and drop ones that are too short to be words.

A waveform with the speech regions highlighted and the silence and noise between them dimmed
Voice activity detection marks where speech is, not what was said

Detection is useful for skipping silence and deciding when you've finished a phrase. With hold-to-talk, as in SayVelio, you decide where the recording starts and ends yourself: it runs exactly while you hold the dictation key.

Why cleaner audio isn't always better

It's tempting to think more noise removal always helps. It doesn't. Consonants like "s", "f", "t" and "k" are quiet, noise-like sounds, and aggressive noise suppression can thin them out along with the background. The audio sounds cleaner to a human, but the recognizer has lost the cues that separate "fifteen" from "fifty".

Modern recognition models are trained on a lot of real-world, noisy speech, so they often cope better with the original sound than with heavily processed audio. Good processing is gentle and is tested on real dictations, not just judged by ear.

Spectrograms: how machines look at sound

Many recognizers don't use the raw waveform directly. They slide a short window (typically 20 to 30 ms) along the audio and measure which frequencies are present in each window. The result is a spectrogram: a picture with time going across, pitch going up, and brightness showing energy.

A spectrogram: time runs left to right, frequency runs bottom to top, brighter means louder
A spectrogram shows how the energy of each frequency changes over time

A spectrogram is not a transcript. It shows patterns of sound that a model learns to connect with speech sounds. The window size is a trade-off: short windows show timing precisely, long windows show pitch precisely. Some newer models skip this step and learn directly from the samples.

What you can do

  • Dictate in the quietest spot you have. Even good processing can't remove a colleague talking next to you, because their voice is speech too.
  • Turn down music and videos while you dictate.
  • Keep a steady distance from the microphone, so the level stays even. More tips in how to dictate well.

Questions

Does better audio quality always produce better text?

Not always. Clear speech at a good level helps the most. Heavy noise reduction, distortion or a wrong audio format can remove details the recognizer needs, and the model's training, language and vocabulary matter as well.

Can dictation work in a noisy café?

Often yes, especially with a headset microphone close to your mouth. Steady noise like a coffee machine is easier to handle than other people's voices nearby.