Can Whisper transcribe in real time?

Yes, with a workaround. The open-source Whisper models that you can run yourself do not stream. They take a finished piece of audio and return text, so a live setup has to cut the microphone signal into pieces, transcribe each piece and show the result as it arrives. The hosted streaming model from OpenAI does this on its own servers. Either way, the words you see lag behind your speech, and some of them change once more audio arrives.

Whisper was built for 30-second chunks, not a live stream

The Whisper paper says its models are trained on 30-second chunks of audio and cannot take longer audio as a single input ¹. That limit is the root of the problem for live use. A sentence that runs past the end of a window cannot be finished until the next window is decoded, so the model works on a window of recent speech, not the full stream.

Streaming recognizers avoid this by returning results as audio arrives. For background on the difference, see how real-time speech recognition works. Whisper is a batch design, so any real-time version is an add-on.

Open-source Whisper streams only through a wrapper that re-runs it

A common route is the streaming example in whisper.cpp, a C++ port of Whisper. Its stream example samples the microphone "every half a second" and runs transcription continuously ². Each pass sends the model a window of recent audio, so the same words are decoded several times, and each new pass can change words from the one before.

The stream example also has a sliding-window mode. Setting its step argument to 0 switches it to transcribing after voice activity is detected ³, so it decodes after voice activity is detected, not on a fixed schedule.

Speed depends on the computer. On Apple silicon, whisper.cpp can run the encoder on the Apple Neural Engine through Core ML, and its project says this can be more than three times faster than CPU-only execution ⁴. A faster machine lets you use a larger model with less delay. On a slow machine, the model falls behind the speaker and the text trails further and further.

The batch throughput of Whisper is also high. On the Open ASR Leaderboard, Whisper Large v3 had an inverse real-time factor of 146 ⁵. That number measures how fast the model works through long recordings, not how soon a word appears on screen, so it does not tell you whether a live wrapper will keep up. For the difference between the two measures, see the glossary entry on real-time factor.

The hosted OpenAI model streams natively, with a different trade-off

OpenAI offers a hosted model called gpt-realtime-whisper for live transcription. OpenAI lists it at $0.017 per minute ⁶. Its turn handling is different: OpenAI says the model needs turn detection omitted or set to null ⁷. Your app, not the service, then decides when a turn has ended, which is the job described in the question on how speech recognition knows when you stop talking.

In an independent 2026 streaming benchmark from Artificial Analysis, the hosted model's word error rate fell from 7.5 percent on its first partial transcript to 5.1 percent on the final transcript ⁸. That is a large gain from waiting. It also shows the trade-off: the first words on screen are less accurate than the words you get when the model settles. The benchmark is a different product from the open-source models, so its numbers do not describe Whisper run on your own computer.

It depends on your hardware, your model size and the delay you accept

Whether Whisper keeps up is a question of three things:

• Hardware. A laptop with a fast neural engine or a graphics card can run a mid-sized model in real time. An older machine often cannot.
• Model size. Smaller models are faster and less accurate. The question on which Whisper model to use covers the sizes in more detail.
• Delay. A window of several seconds adds that much delay before the text is final. Shorter windows respond faster but give the model less context, which can cost accuracy.

For clean English in a quiet room, a small or medium model on decent hardware can be good enough for captions. For noisy rooms, accents or specialist vocabulary, the larger models are more accurate, but they need more power. The guide on running Whisper locally sets out the setup steps for each platform.

What to do if you need live text

1. Decide what live means. If you need captions that flicker and catch up, a whisper.cpp wrapper or a similar tool is enough. If you need correct text soon after you stop, batch the final pass and show the partial text only as feedback.
2. Test before you commit. Run your own recordings through the tool you pick. Measure the time from when you stop speaking to the final text, and count the errors with word error rate. A fast result with many errors costs more time to fix than a slightly slower correct one.
3. Check the hardware first. On a Mac with Apple silicon, the Core ML path is the one to try first. On other machines, start with a small model and move up only if the delay is acceptable.
4. Use a ready-made app if you do not want to build a pipeline. Yap, which we make, runs on-device dictation on Mac and Windows, so you can dictate without setting up a streaming wrapper. See the download page for the current version, and the Windows page for the Windows setup.

Whisper can transcribe in real time, but only as a wrapper around a model that was designed for batches, or through a hosted model whose turn handling you must manage. Choose the one whose delay and errors suit your use, and test it on your own audio before relying on it.

For the speed question across services, read the question on the fastest speech-to-text API.

For the wider field, read our overview of speech recognition.