Self-hosted realtime speech-to-text for voice agents

Every voice agent hears, thinks and speaks. lini is the hearing part: realtime speech-to-text that runs on your own servers, with the open-source model you choose.

Why teams choose self-hosted speech-to-text

• Privacy: callers' audio stays inside your infrastructure. Nothing is sent to us.
• Cost: you pay for servers and a licence, not for every minute your bot listens.
• Choice: pick the model that fits your language and accuracy needs, and switch when a better one comes out.
• Control: no third-party rate limits or outages in the middle of a call.

The problem with open-source models

The most accurate open models, like Whisper and Parakeet, are offline. They wait for the full sentence, so callers sit in silence. That silence is when people say "Hello? Are you there?" and hang up.

What lini changes

• Words arrive while the caller is speaking
• Final text lands soon after they stop, so your agent can reply quickly
• Works with phone audio (8 kHz) and WebRTC (16 or 48 kHz)
• CPU or GPU, in your cloud or on-premises
• No retraining and no new data

How it fits your stack

lini sits between your audio and your model, and hands live text to the rest of your agent: your language model, your voice, your telephony. On a call we look at your stack together (custom code or popular voice-agent frameworks) and agree how to connect it.