Cartesia Ink vs Deepgram Speech-to-Text (Nova-3, Flux)

Cartesia Ink

Ink 2 streams transcripts with turn detection built into the model, documented by an AsyncAPI file with typed ranges and structured errors. The terms, last revised 14 June 2024, let Cartesia train on inputs and outputs unless the customer opts out, and zero data retention is an Enterprise setting. Ink 2 has no batch endpoint and no diarisation.

Deepgram Speech-to-Text (Nova-3, Flux)

Flux streams with model-level end-of-turn detection, so a voice agent needs no separate VAD. Training on audio is the default and the opt-out is a per-request flag.

Cartesia Ink

1. Use wss://api.cartesia.ai/stt/turns/websocket with model=ink-2, encoding, sample_rate and cartesia_version=2026-08-14. All four are required.
2. Send raw mono audio in chunks of about 100 ms at the speed it was spoken. Pushing a whole file into the socket can return an internal server error.
3. Read the final text from turn.end only. transcript is cumulative within a turn, so joining turn.update events duplicates text.
4. Send {"type": "close"} after the last audio and keep reading until the server closes the socket, or the buffered tail is lost.
5. Check encoding and sample_rate against the source before sending. The docs say the server might not return an error when they are wrong.

Deepgram Speech-to-Text (Nova-3, Flux)

1. Add mip_opt_out=true to every request that carries customer audio
2. Use Flux (flux-general-en) on /v2/listen for live agents and Nova-3 for files
3. Back off exponentially on 429. The concurrency limit is per project
4. Pass callback for long files so the request doesn't hit the 10-minute processing timeout
5. Mint keys with an expiry for short-lived jobs