Skip to main content
Ink 2 is our fastest, most accurate multilingual streaming speech-to-text model — the lowest word error rate and best turn detection of any streaming STT model, specifically built for production voice agents. It transcribes structured data like phone numbers, dates, and emails right the first time, and knows when a speaker starts and finishes from the semantics of the conversation. Turn detection is built in, so you don’t need a separate voice activity detector (VAD). Ink 2 sends turn events (turn.start, turn.update, turn.eager_end, turn.resume, and turn.end), so your agent knows exactly when to listen and when to respond. See Turn detection for how the events work, or Compare STT endpoints to run Ink 2 with or without turn detection. ink-preview runs our newest multilingual model. It supports 44 languages, returns word timestamps and detected languages, takes language hints, and lets you adjust keyterms and hints mid-stream. It can change without notice, so use ink-2 for production traffic that needs stable behavior. For information on ink-whisper, see our page on Older STT Models.

ink-preview - November 2026

New Features:
  • More languages. Supports 44 languages. Add language hints to steer transcription toward the languages you expect.
  • Language detection. Returns a list of languages detected and their confidence scores.
  • Word timestamps. Returns a start and end time for every word.
  • More keyterms. Up to 500 keyterms totaling 5,000 characters, and you can change them mid-stream.
Both ink-2 and ink-preview can end a turn with an empty transcript. Skip these turns instead of replying. See Turn detection.

Where to go next

Try it out online

No sign-up or code required

Start building

Guides and best practices