Voice Cloning And Voice Library

Operator uploads a sample clip to create a cloned voice and expects an instantly usable, high-fidelity voice id from a 2-second noisy phone snippet.

Create the voice via the documented voice-creation request with an adequate, clean sample (per documented clip duration/quality guidance [REQUIRES-VERIFICATION on exact minimums]); persist the returned voice id with its provenance. Validate input audio quality before submission rather than expectin…

After cloning, the agent re-uploads the source clip on every /tts call instead of referencing the persisted voice id.

Reference the persisted voice id in subsequent /tts requests. Do not re-upload or re-clone per call — that wastes quota and may produce subtly different voices. Keep a stable id→voice mapping in the operator's store.

Agent mixes a request that specifies voice by id with a hand-edited embedding array in the same field, expecting a blend.

Specify a voice either by its id or by a voice embedding object per the schema — not a malformed hybrid. If using an embedding (e.g. for mixing/customization), use the documented embedding shape and obtain it through the supported path, not by hand-mutating array values.