Kalpa Labs

Developers who want to add natural-sounding speech to products today must pick between TTS systems from providers like Cartesia, ElevenLabs and OpenAI, which Kalpa Labs' own blind listening benchmark suggests still sound less human than its model. Most existing audio models are also narrow, built to do one task such as plain text-to-speech, rather than following open-ended instructions the way modern text LLMs do. Teams building for markets like India also need natural code-switching between English and Hindi, including English words inside Hindi sentences, which is not a given in most TTS systems.

Kalpa Labs offers kalpa-tts-1, a text-to-speech API available now that generates speech in eleven voices across English and Hindi, including code-switched sentences, with about 300ms to first audio over a streamed WebSocket connection returning base64 pcm_s16le or 24kHz mono 16-bit WAV. It includes 1,000 free minutes for three months before moving to a contract. Beyond shipped TTS, the company is training a base audio model (an 8B backbone with a 300M decoder, interleaved text and audio, 32 codebooks) from random weights on millions of hours of real, weakly labelled in-the-wild audio, aiming toward a single model promptable with both audio and text for tasks like continuing a scene, cloning a voice's delivery, or changing speaking style, currently previewed and gated behind a waitlist.