Realtime Whisper, on your own servers

OpenAI Whisper is accurate and speaks almost a hundred languages. It was built for finished recordings, though. lini lets you use the same Whisper model in live products, with words on screen while people are still talking.

Why Whisper isn't realtime out of the box

Whisper reads audio in windows of up to 30 seconds and writes the text when the window is done. That's perfect for a podcast or a recorded meeting, and too slow for a person waiting on a reply.

The quick fix people try is to cut the audio into one-second pieces. In our tests that made Whisper repeat words, invent words that were never said, and cut words in half at the edges. It also ran slower than realtime on an ordinary CPU.

What changes with lini

• Words appear while the person is speaking, then settle into final text.
• Final text arrives soon after the person stops, so your app can act on it.
• The Whisper weights stay exactly as they are. No retraining, no new data.
• It runs next to your model, on your servers. Audio never leaves your setup.

How fast is realtime Whisper?

It depends on the Whisper size and your hardware. Large Whisper models are heavy on a CPU, so for phone bots that run on CPU only we usually suggest a smaller Whisper or a GPU. If you need English or European languages on plain CPUs, NVIDIA Parakeet gave us final text in about half a second to a second after the speaker stopped.

On a call we measure speed and accuracy on your own recordings before you decide anything.

Where realtime Whisper fits

• Live captions for meetings, classes and events in many languages
• Self-hosted voice agents where audio must stay in-house
• Calls in languages that English-only models don't cover, including Hindi
• Teams that already run Whisper for batch work and now want it live