VAANI by Bibha AI — Voice Analysis And Natural Interpretation — is Bibha’s self-hosted text-to-speech (TTS) model family. VAANI voices every live voice agent and every Voice Studio generation, running entirely on your organization’s own GPU infrastructure: no request ever leaves for an external speech provider.

​
The VAANI family

All four variants support voice cloning: upload or record a short reference sample and VAANI speaks in that voice. VAANI additionally understands paralinguistic tags embedded in the text — [clear throat], [sigh], [shush], [cough], [groan], [sniff], [gasp], [chuckle], [laugh] — so an agent can chuckle or sigh mid-sentence. The other variants have no tag tokens and would speak them literally.

​
Capabilities

  • Voice cloning — clone a voice from seconds of reference audio; profiles are managed on the Voices tab or via the Voice Profiles API.
  • Emotion and style control — exaggeration, pacing (CFG), temperature, repetition penalty, and nucleus/top-k sampling, per variant (surfaced in the Studio’s Fine-tune panel). These are model-level knobs; which ones a variant honors is advertised per engine, so a control is only shown when it actually does something.
  • Post-processing — speed (pitch-preserving time-stretch), pitch shift, volume, sample rate, and output container are applied to the rendered audio after synthesis, so they work identically on every variant. No VAANI engine takes a pitch or speed parameter natively.
  • Streaming synthesis — low time-to-first-audio for live calls, with incremental PCM streaming on the OpenAI-compatible POST /v1/audio/speech route.
  • Self-hosted — served from your organization’s own GPU plane, metered per organization, with no external speech provider in the path.

​
Using VAANI

VAANI is selected like any engine in the Studio’s Generation box or an agent’s voice configuration. Programmatic synthesis goes through the Speech Generation API and the realtime conversation routes.
Engine identifiers used before the VAANI catalog was consolidated are still accepted and resolve to the variant above, so existing saved agent configurations and voice profiles keep working unchanged. Use the identifiers in the table for new integrations — they are the values the API returns. A profile still recorded against an engine that has since been retired reads back as unavailable; re-point it at one of the four above.