Speech to Text Glossary: 20 Terms Explained Simply

Speech to text is full of jargon: ASR, WER, VAD, diarization, on-device, streaming. Most of it sounds harder than it is. This glossary defines 20 core terms in plain English so you can compare voice to text and dictation tools on Mac without a computer science degree.

Speech to text, in plain words

Every voice typing app, from Apple's built-in dictation to a dedicated Mac tool like BlaBlaType, is built on the same building blocks. Vendors describe those blocks with a shared vocabulary, and once you know it, spec sheets stop being intimidating. Most people speak around three to four times faster than they type, so it is worth learning enough of the language to pick a tool that actually keeps up.

If you want the wider picture first, our complete 2026 guide to voice to text on Mac walks through setup end to end. This page is the dictionary you can keep open beside it.

The 20 terms explained simply

• 1Speech to text (STT). The umbrella technology that converts spoken audio into written words. Everything else here is a piece of it.
• 2Dictation. Using speech to text live, so words appear where your cursor is as you talk. It is the everyday, hands-free way to write.
• 3Transcription. Turning a recording, such as a meeting or voice memo, into text after the fact rather than live.
• 4Automatic Speech Recognition (ASR). The engine that actually maps sounds to words. When people say "the model," they usually mean the ASR system.
• 5On-device processing. The recognition runs on your own Mac, so audio is converted to text locally and never uploaded.
• 6Cloud processing. Your audio is sent to a remote server, transcribed there, and the text is sent back. Fast, but your voice leaves the device.
• 7Whisper. A widely used open speech recognition model family, strong across many languages, and able to run fully offline on a Mac.
• 8Parakeet. Another modern ASR model line, known for speed and low latency, and also usable on-device.
• 9Voice Activity Detection (VAD). The step that decides when you are actually speaking, so silence and background noise are ignored.
• 10Word Error Rate (WER). The standard accuracy score. It counts substituted, inserted, and deleted words over the total spoken. Lower is better.
• 11Latency. The delay between speaking and seeing text. Low latency feels instant; high latency feels laggy.
• 12Filler words. The "um," "uh," and "you know" that pad natural speech. Good tools strip them out automatically.
• 13AI cleanup (post-processing). A second pass that fixes punctuation and grammar, removes filler, and can adjust tone, turning raw speech into polished text.
• 14Custom dictionary. A personal word list for names, brands, and jargon so the model spells them right every time.
• 15Speaker diarization. Labeling who said what in a recording with more than one voice. Useful for meeting transcripts.
• 16Language model. The part that predicts likely word sequences, so "recognize speech" wins over "wreck a nice beach."
• 17Punctuation restoration. Adding commas, periods, and capitals that you did not literally say aloud, so the text reads naturally.
• 18Push-to-talk. A shortcut you hold to dictate and release to stop, versus a toggle that stays on until you switch it off.
• 19Real-time streaming. Transcribing continuously as you speak instead of waiting for you to finish a whole clip.
• 20Translation (translate-as-you-speak). Speaking one language and getting text out in another, live, on top of plain transcription.

Ready to put the vocabulary to work? Our walkthrough on building a voice-first writing setup on Mac shows how push-to-talk, a custom dictionary, and AI cleanup fit together in a daily workflow.

On-device versus cloud: the terms that matter for privacy

If only one distinction on this list sticks, make it this one. Where the recognition runs decides where your voice goes. Here is the honest trade-off.

BlaBlaType sits firmly on the on-device side: speech recognition runs 100% locally on Apple Silicon, so nothing is uploaded. If you are comparing tools on this exact point, our offline Wispr Flow alternative breakdown puts cloud and local dictation side by side. For a sense of how far open models have come, the original Whisper research paper is a readable starting point.

Myth versus fact

A few of these terms get misused in marketing. Three quick corrections.

Putting the glossary to work on your Mac

Once the terms click, choosing a tool is mostly about matching features to the words: do you need on-device processing, AI cleanup, a custom dictionary, and translate-as-you-speak, all system-wide in any app? On Mac, that combination is exactly what BlaBlaType offers, and there is a three-day free trial with no card so you can test the accuracy on your own voice. Teams who write repetitive text should also read our note on how to dictate documentation and SOPs by voice. You can compare tiers any time on the pricing page. For context on how conversational voice features work in other tools, OpenAI's voice mode FAQ is a useful reference.