Speech-to-text

Audio capture turns sound into a stream of samples. Feature extraction converts those samples into a representation the model can read, usually a spectrogram. The recognition model maps that representation to text, sometimes through an intermediate step of phonemes or sub-word tokens. Post-processing turns the model’s output into something a person would write: punctuation, capitalization, numbers as digits, and in the better tools, filler removal and tone.

When people say a tool "uses AI", they almost always mean the third stage. The fourth is where most of the daily-use quality lives, and it is the stage the marketing tends to skip.

Cloud STT sends audio to a server, runs a large model there, and returns text. It can use models too big for a laptop and it works on any device with a network. On-device STT runs the model on your own hardware, so audio never leaves the machine, it works offline, and there is no round trip. On Apple silicon the gap in accuracy has narrowed to the point where the choice is about privacy and cost far more than quality.

"Speech-to-text" on a spec sheet usually means the raw model. "Dictation" means the whole experience, including where the text lands. "Transcription" means files. "Streaming" means words appear as you speak rather than after. "Local" or "on-device" means the audio stays home. Knowing which stage each word describes is most of what you need to read a comparison honestly.

Every Mac has speech-to-text built in. Apple calls it Dictation, and it types into any text field where the cursor is. You can set it up in about a minute.

1. 01 Choose Apple menu → System Settings, then click Keyboard in the sidebar.
2. 02 Go to Dictation and turn it on.
3. 03 Click the pop-up menu next to Microphone source and choose the microphone you want. Automatic lets the Mac listen to the device you are most likely to use.
4. 04 Click the Shortcut pop-up menu and choose a shortcut to start Dictation. Choose Customize to make one that is not in the list.
5. 05 Click in any text field, press the shortcut, and speak.
6. 06 Stop when you are done. Dictation also stops automatically when no speech is detected for 30 seconds.

It depends on the audio more than most people expect. NVIDIA publishes results for its open Parakeet TDT 0.6B v3 model on eight English test sets, and the same model scores a word error rate of 1.93% on clean read speech (LibriSpeech test-clean) and 11.31% on meeting recordings (AMI).

Noise moves the number too. On the same model card, the average word error rate rises from 6.34% on clean audio to 11.66% when noise is as loud as the speech (a signal-to-noise ratio of 0). A quiet room and a close microphone are the cheapest accuracy upgrade you can make.

• Kind of speech: read sentences are easier than a meeting with people talking over each other.
• Background noise and music: error rates climb as the noise gets closer to the level of your voice.
• Accent and dialect: the Whisper model card reports different performance across accents and dialects.
• Language: the Whisper model card reports lower accuracy on languages with less training data.
• Vocabulary: the Parakeet model card warns that a word outside the model’s vocabulary is not likely to be recognized.

Speech-to-text is either processed on your device or sent to a server, and you can usually find out which. For Apple’s Dictation the answer is written in Settings. For any other tool, the permission list tells you which apps can hear the microphone at all.

1. 01 Choose Apple menu → System Settings, then click Keyboard in the sidebar.
2. 02 Go to Dictation and read the text under the Dictation switch. It says whether your voice input and transcripts are processed on your device.
3. 03 Go to System Settings → Privacy & Security → Microphone to see every app that has asked to use the microphone.
4. 04 Turn off any app in that list that you do not want listening.
5. 05 For a third-party tool, read its privacy policy for where audio is processed. The Microphone list shows who can listen, not where the audio goes.