Build with Ink

The quickest way to get started with Ink is by using an integration. If your preferred integration isn’t listed and you already use another speech-to-text model, the migration guides explain how to wire Ink into an existing codebase. See the references section for complete documentation of all API endpoints and best practices.

​

Integrations

• Line
• LiveKit
• Pipecat

​

Comparing Ink against other models

Ink supports two realtime transcription modes:

1. Client sends audio (Auto finalization)
2. Client sends audio and signals when to finalize transcripts (Manual finalization)

Most speech-to-text APIs combine both behaviors, but Cartesia separates them for improved model performance. Auto finalization is recommended for most agents. At the same time, there are many use-cases where manual finalization is necessary. Some examples are:

1. Push-to-talk apps
2. Pipelines where you know speech is over and are waiting for the transcript

​

Migration guides

• Deepgram Turn-based Audio (Flux)
• Deepgram Live Audio (Nova)
  • Auto finalization
    Ink automatically finalizes transcripts
  • Manual finalization
    Your client decides when to finalize transcripts
• ElevenLabs Realtime Speech to Text
  • Auto finalization
    Similar to ElevenLabs’s commit_strategy=vad
  • Manual finalization
    Similar to ElevenLabs’s commit_strategy=manual
• OpenAI Realtime Transcription
  • Auto finalization
    Similar to OpenAI’s turn_detection: server_vad
  • Manual finalization
    Similar to OpenAI’s turn_detection: null
• OpenAI Speech to Text
  Batch audio transcription

​

References