Realtime NVIDIA Parakeet, without retraining

NVIDIA Parakeet is one of the most accurate open speech-to-text models, and it's light enough for ordinary CPUs. It's an offline model, so it expects finished audio. lini turns the Parakeet you already have into a live stream.

What you get

• Words on screen while the person is still speaking
• Final text about half a second to a second after they stop, in our tests on an ordinary CPU
• Phone audio (8 kHz), web and mobile (16 kHz) and studio microphones (48 kHz)
• The same Parakeet weights. No retraining, no streaming re-design

Why not train a streaming model instead?

Turning a model into a streaming one usually means new training runs, new data and weeks of work for every language. You also end up maintaining a second model. lini skips that: it makes the model you already trust live, so when a better open model comes out you can switch to it.

Numbers, not just a claim

Parakeet got every digit right in our phone-quality tests of account numbers and amounts, which matters for banking and support calls. On real earnings-call recordings, the live text from lini stayed within about a point of Parakeet's accuracy on the whole recording.

Your audio is different, so on a call we run lini on your recordings and show you the numbers.

Good fits for realtime Parakeet

• Voice agents and phone bots that need fast replies on CPU
• Contact-centre agent assist, with notes while the customer talks
• English and European-language captions
• Teams moving off per-minute cloud speech-to-text