Kalpa Labs is an audio research lab working towards generalist audio models: models that follow instructions and learn new tasks in context. Our first conversational speech models are in beta today. Direct a scene in Studio, talk to a model live, or build on the API.
Write the lines, cast the voices, set the mood. Studio performs the whole scene: multi-speaker dialogue with pacing, emotion, and voices that stay themselves.
Open Studio→Open a call in your browser and just talk. Ask it anything and hear it answer in its own voice. The beta model, live and unedited.
Say Hello→The same models over a clean REST API. Docs written for humans and for models, where every page has a markdown twin, plus a playground that runs in the browser.
Read the Docs→Speech models today are roughly where language models were a few years ago: they can speak, but they can't really listen. They don't follow instructions about sound, learn a voice or a style from a few examples in context, or notice how something was said and not just what. Generalist audio models that close that gap are what this lab exists for.
So we train our models end to end, from data and tokenizers through pretraining, post-training, and inference, because the capabilities we're after don't come from fine-tuning someone else's checkpoint. The conversational models in beta are the first public step. We'll make bigger claims when we can demo them.
Until then, the beta is open. Go talk to it→Training runs, data pipelines, inference engines, and the products on top, all built in-house by a team you can count on two hands. If you'd rather ship the model than ship around it, we should talk.