Two model tiers. Both cover the whole voice catalog; they differ in fidelity, speed, and price.

Sada — sada-1

Flagship expressive model. Highest fidelity, full emotional range. The default.

Nabra — nabra-1

Fast and economical. Built for high-volume production traffic.

​
Comparison

​
Choosing

Use sada-1 for anything a customer hears in a live conversation. It is the default, it is the only tier with a streaming path, and expressiveness matters most exactly where latency does. Use nabra-1 for bulk generation where cost dominates and nothing is waiting on the result — pre-rendering prompt libraries, generating notification audio in batch, non-realtime pipelines. At 2 cents per 1,000 characters it is 60% cheaper.
nabra-1 cannot stream. /v1/speech/stream and the WebSocket accept sada-1 only, and reject other models up front rather than failing mid-stream.

​
Pitch

pitch applies only to nabra-1. The flagship tier ignores it — its expressiveness comes from the model rather than from post-hoc adjustment, and a pitch-shifted flagship voice would sound worse, not more configurable.
Use speed for pacing on either tier.

​
Cost

Cost is per 1,000 characters, rounded up, so a very short request still costs at least one cent. Every synthesis response returns X-Voho-Characters and X-Voho-Cost-Cents, so you can attribute spend per request without a separate usage call.
Rounding up means many tiny requests cost more than the same text batched. If you are generating a prompt library, synthesise whole prompts rather than fragments.