Ink 2 Multilingual

Cartesia's ink-2 for callers speaking English, French, Hindi, Japanese or Spanish. It detects the language itself on every turn, and decides the end of a turn on its own semantic judgement rather than a silence timer, like the English-only ink-2.

Reached through Cartesia's own API with OnCore's Cartesia key, not the LiveKit Inference gateway.

The same ink-2 model as the English entry, reached by a route that lets it hear all of its languages.

Settings#

Every setting is optional. Leave one out and OnCore sends nothing for it, so the Default applies. If invalid is what the provider does with a value it does not accept.

Overriding on a single call#

Return this from your assistant-request webhook to change these settings for one call only. The stored transcriber is left untouched, and the next call starts from it again.

{
  "data": {
    "overrides": {
      "transcriber": {
        "settings": {
          "turnStartThreshold": 0.7,
          "turnEagerEndThreshold": 0.5,
          "turnEndThreshold": 0.2,
          "turnEndTimeoutMs": 2000
        }
      }
    }
  }
}

Send only the keys you want to change: anything you leave out keeps the value the assistant already has. Add "id" beside "settings" to swap the whole transcriber for another one from your catalogue at the same time.

turnStartThreshold: How sure the model must be that the caller has started speaking before it opens a turn. Must stay above turnEagerEndThreshold.

turnEagerEndThreshold: The level below which the model signals that the caller may be finishing. Must sit between turnEndThreshold and turnStartThreshold.

turnEndThreshold: The level below which the model ends the turn and the assistant replies. Lower waits for more certainty that the caller has finished. Must stay below turnEagerEndThreshold.

turnEndTimeoutMs: The longest the model waits after the caller stops speaking before ending the turn anyway, however unsure it still is.

Languages#

This transcriber covers the five languages Cartesia publishes for ink-2: English, French, Hindi, Japanese and Spanish. There is no language to set. The model works out which of the five is being spoken on every turn, so a caller who switches from English to Spanish part way through is transcribed correctly in both. A language on the transcriber is not sent to Cartesia.

The assistant settles on one of the five over the caller's first three turns, and its fixed lines, such as the goodbye, are spoken in that language from then on. A caller who changes language after that is still transcribed correctly, but the fixed lines stay in the language the call settled on. See multi-language calls.

For a language outside these five, use ink-whisper.

Turn-taking#

Like the English ink-2, this model judges the end of a turn on its own rather than waiting out a silence timer. The four settings above tune that judgement. The three thresholds have to stay in order: turnStartThreshold above turnEagerEndThreshold, which is above turnEndThreshold. A threshold you leave out still counts at its default, so setting turnStartThreshold to 0.55 on its own breaks the order against the default 0.6 of turnEagerEndThreshold.

A value Cartesia would refuse is dropped rather than sent. Cartesia ends the session on a value outside its range or on thresholds out of order, which would leave the call with no transcriber. OnCore checks every value before connecting and drops the ones Cartesia would refuse: an out-of-range value on its own, and all three thresholds when their order is wrong. Cartesia's default applies in place of anything dropped, and the call goes ahead.

Recognition terms#

Terms from a dictionary attached to the assistant are sent to Cartesia on every call. They make the names and words in them more likely to be written correctly, though not certain to be: measured in English, French and Spanish, terms corrected most of the names they listed, not every one. Cartesia accepts up to 100 terms totalling 1200 characters. OnCore sends terms in order until the next one would pass either limit, and leaves out the rest.

What this transcriber does not take#

minVolume and maxSilenceDurationSecs belong to ink-whisper and are not sent from here. extraKwargs are not carried on this route either: the four settings in the table and the dictionary's terms are all that reach Cartesia.

Next steps#

• Ink 2 and Ink Whisper: the previous Cartesia model.
• xAI: speech-to-text from xAI.
• Providers introduction: per-call overrides, and how a wrong value behaves across providers.