Azure AI Speech speech-to-text vs Cartesia Ink

Azure AI Speech speech-to-text

Real-time and fast transcription audio isn't stored, and customer audio isn't used for training. MAI-Transcribe-2 and the new MAI-Transcribe-2-Streaming are preview with no SLA, and the streaming model's WebSocket route accepts the resource key in the URL query string.

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.

Azure AI Speech speech-to-text

1. Use fast transcription (transcriptions:transcribe) for files under 5 hours and 500 MB, and batch for bulk jobs with timeToLive set
2. Pin api-version=2025-10-15. v3.0 and the v3.2 previews are retired
3. On a 429, back off 1, 2, 4 then 4 minutes. It usually means autoscaling, not a quota
4. For MAI-Transcribe-2-Streaming, use Speech SDK 1.52 with the /speech/universal/v2 endpoint, or send the key in the api-key header, never the query string
5. Don't budget on MAI-Transcribe-2 at $0.10 or MAI-Transcribe-2-Streaming at $0.54 an hour after 2026-12-31

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.