Most natural real-time TTS with voice cloning and sub-60ms latency, on-prem or via API. Grammar-aware normalization reads phone numbers, IBANs, addresses, and medications naturally across 25+ languages, with word-level timestamps and IPA support. Adapters for LiveKit, Pipecat, and Vapi. Built by 4 in Berlin.

Hi PH 👋 Kajo from KugelAudio here. We're 4 people in Berlin, currently in SF for YC. Building for the future, which we believe will be conversational. Today we're shipping a real-time TTS model with voice cloning. If you only have a minute, here's what I'd want you to know. You can clone a voice from 30 to 60 seconds of audio. Drop in a short sample and you get a working voice immediately. We optimized for voice agents with latencies below 60ms (excl. network), input streaming and output streaming. We offer on premise support, run the model in your own cluster instead of calling our API. Useful when data needs to stay on your network. Adapters for LiveKit, Pipecat, and Vapi. SDKs in Python, JS, and Java. Free tier so you don't have to talk to us before trying it. Alex (our founding engineer) is in the thread today for the hard questions: model architecture, where the latency came from, what we broke before this version worked. Ask anything!

Congratulations on the launch guys! Few questions:
1. Websites list languages like Hindi to be supported but can not find any voice related to that on playground.

2. I checked the websocket streaming API for TTS.. is it possible to have multi context support (just like elevenlabs) in the streaming API? is that part of plan in future?

Cheers!

Sensational for German TTS

Congrats on the launch guys!