Sonic Tts Synthesis

Agent calls POST /tts/bytes with transcript and voice but omits model_id and output_format, assuming server defaults.

Send model_id (e.g. a Sonic model id), transcript, voice (id or embedding), and a fully-specified output_format object (container, encoding, sample_rate). Do not rely on implicit defaults for format — the byte stream's decodability depends on the requested container/encoding being explicit.

Operator hardcodes a Sonic model_id captured from a blog post a year ago and never revisits it across releases.

Pin model_id explicitly per docs, but treat the id as a managed dependency: track model availability/deprecation from docs.cartesia.ai and have a tested upgrade path. Do not silently fall back to an unspecified default if the pinned id is retired — surface the error and re-pin deliberately.

Agent needs a single short confirmation phrase synthesized for a non-interactive batch report, but wires it through /tts/sse and reassembles chunks.

For a complete, non-streamed clip use POST /tts/bytes and consume the full audio body. Reserve /tts/sse (Server-Sent Events) for cases that benefit from incremental playback. Match the endpoint to the consumption pattern instead of over-engineering chunk reassembly for a one-shot clip.