We’re building one model for all of audio.
It starts with speech. Try kalpa-tts-1 in English and Hindi.
State-of-the-art audio realism.
Expert listeners hear two clips, blind and in random order, and pick the one that sounds more human. Here is Kalpa TTS Beta V0.1 next to the best models from Cartesia, ElevenLabs and OpenAI, and Eleven v4.
Elo from AudioRealismBench by Design Arena, updated 11 Oct 2026.
kalpa-tts-1, in the API.
Text-to-speech in eleven voices: English and Hindi, including English words inside Hindi sentences.
curl https://api.kalpalabs.ai/v1/tts/5e0c5704-590f-483b-b291-00a2415cb67e \ -H "Authorization: Bearer $KALPA_API_KEY" \ -H "Content-Type: application/json" \ -d '{"text": "नमस्ते! आज आपके लिए क्या पढ़ूँ? कोई story, कोई news, या आपका अपना message?"}'
One model you prompt with audio and text.
Text models went from continuing text to following instructions. Most audio models still do one task each. Our base model, trained from random weights on millions of hours of real audio, already takes the first steps.
Broadcast this like 1938 news radioStyle prompt
Speak like sports commentarySame sentence, another prompt
Towards generalist audio models
The architecture, the data, a blind preference eval against four systems, and what the base model can already do.
Read the reportInterleaved text and audio; 32 codebooks, 31 of them from the RVQ decoder.
No text LLM underneath.
Weakly labelled real transcripts; ASR-like ones filtered out.
The room is part of who is speaking.
A proposed eval: how long can a person talk with a model before deducing, reliably better than 50%, that it is an AI?