Streaming transcription
Also: real-time transcription
Updated
Definition
Streaming transcription shows words on screen while you are still speaking, rather than after you stop. The alternative, batch transcription, processes the whole utterance at once and shows the text at the end.
Streaming requires a model built to emit partial results and revise them as more audio arrives, which is a different engineering problem from transcribing a finished recording. Not every ASR model can stream; products often stream in some languages and fall back to batch in others.
For dictation, streaming is mostly about confidence: seeing the words arrive confirms the tool heard you. The final text quality is decided by what happens at the end of the utterance either way.
Why streaming makes dictation feel instant
With batch recognition, the model starts working when you release the key and you wait for the whole recording to be processed. The pause is proportional to how long you spoke. With streaming, the model has been keeping up all along, so at key-up there is only the last fraction of a second of audio to finish and the formatting pass to run. The result is a delay that does not grow with the sentence, which is what people mean when they say a tool is fast.
There is a second effect. Watching the live words tells you the tool is hearing you correctly, so you stop glancing at the result afterward to check. That trust is most of what makes dictation a habit rather than an experiment.
Partial results and revisions
A streaming model commits words tentatively and may change its mind as more audio arrives: "the meeting is" can become "the meeting isn’t" a moment later. Good tools show these revisions in a transient display, like a notch overlay or a floating pill, and only insert the settled text into your document. Inserting live partials into the app and then rewriting them is how a tool ends up fighting with your text editor.
Why not every language streams
Streaming needs a transducer-style model trained for the language. The families that stream well today cover a few dozen languages; batch models like Whisper cover about a hundred. A tool that claims 99 languages and live text is usually streaming in the languages its transducer supports and batch-processing the rest, and an honest product page says which is which.
How to see streaming transcription on your Mac
macOS has a built-in example of streaming transcription called Live Captions. It writes captions for audio from your apps, or for sound picked up by the microphone, while the audio is still playing. Apple says the audio is processed on device. Live Captions is available only on Mac computers with Apple silicon, and not in all languages, countries, or regions.
- Choose Apple menu → System Settings, then click Accessibility in the sidebar.
- Click Live Captions and turn it on.
- Play a video or a podcast in any app.
- Watch the captions window. The text appears while the speaker is still talking, which is streaming.
- Turn Live Captions off in the same place when you are done.
Can Whisper and Parakeet transcribe in real time?
Not out of the box, in Whisper’s case. OpenAI’s model card says Whisper models cannot be used for real-time transcription out of the box, and that others may be able to build near-real-time applications on top of them. The reference code moves a 30-second window across a finished file.
Parakeet is closer. NVIDIA’s model card for Parakeet TDT 0.6B v3 documents a streaming mode, and its example settings show the trade every streaming setup makes: audio is processed in 2-second chunks, with 10 seconds of earlier audio and 2 seconds of later audio as context. The later audio is look-ahead. More of it gives the model more to go on, and it also means the words appear later.
Streaming or batch: which one do you need?
Streaming matters when a person is waiting on the words. Batch is fine, and often simpler, when the audio already exists as a file. If you are choosing a tool, match the mode to the job.
- Dictating into a text field: streaming, so you can see that you are being heard.
- Captions for a live call or a talk: streaming, because a caption that arrives after the sentence is over is of little use.
- Transcribing a recorded meeting, interview, or lecture: batch. Nobody is watching the words arrive.
- Making subtitle files: batch. The Whisper command-line tool writes srt and vtt files directly.
- A language your streaming model does not cover: batch, with a model that does cover it.
Questions people ask.
What is the difference between streaming and batch transcription?
Streaming shows words as you speak and finishes almost immediately when you stop. Batch waits until you stop, then processes the whole recording, so the pause grows with the length of what you said. Both can be accurate; streaming feels faster and shows you it is listening.
Does streaming make the text less accurate?
Not in the settled result. A streaming model may revise a word as more audio arrives, but what gets inserted is the final text. The accuracy you experience depends more on the model quality and your microphone than on whether it streamed.
Why does my dictation tool stream in English but not in my other language?
Because its streaming model does not support that language and it is falling back to a batch model. The text should still be right; it will just arrive after you stop rather than while you speak.
Can Whisper do real-time transcription?
Not out of the box. OpenAI’s model card says Whisper models cannot be used for real-time transcription as shipped, though others can build near-real-time applications on top of them.
Does a Mac have built-in real-time transcription?
Yes. Live Captions, under System Settings → Accessibility → Live Captions, shows captions for audio as it plays. It is available only on Mac computers with Apple silicon, and not in every language or region.
Does real-time transcription need an internet connection?
Not necessarily. Streaming describes when the words appear, not where the audio is processed. Apple says Live Captions on a Mac processes audio on device.
