Voice model benchmark

Type text once — it's translated into Hindi, Kannada, and Tamil, then synthesized in all 3 languages across both confirmed-working voice models (Sarvam, Gnani), so you can compare latency and voice quality by ear across every combination. Internal eval tool, not part of the customer-facing call.