Hearing
The speech recognition engine, the call language, and the noise suppression that cleans the caller's audio before the agent hears it.
Hearing is the listening half of a call: the engine that transcribes the caller, the language the call runs in, and how their background noise is cleaned before it arrives.
Two parts of it have pages of their own:
- Turn taking, for how patiently the agent waits before deciding the caller has finished.
- Recognition hints, for the words the agent must not garble.
The settings on this page are versioned. They belong to the draft, reach live calls when you publish, and are read-only on a published version.
Speech to text
The engine that turns the caller's audio into text.
| Engine | Provider | Languages | Turn detection | Recognition hints |
|---|---|---|---|---|
| Ink 2 | Cartesia | English only | On the engine | Yes |
| Nova 3 | Deepgram | Multilingual | Platform | Yes |
| Flux | Deepgram | English only | On the engine | Yes |
| Nova 2 | Deepgram | Multilingual | Platform | Yes |
| Ink Whisper | Cartesia | Multilingual | Platform | No |
Ink 2 is the default
Engines are grouped by provider in the picker, the platform default is marked, and Use platform default returns an agent to it.
Ink 2 is the platform default. It decides turn boundaries on the engine itself, which removes a hop and makes conversations feel tighter. It is English only and accepts recognition hints. An agent that has never picked an engine runs it.
Nova 3 is the right alternative when you need multilingual support.
Who decides the caller has stopped
Nova 3, Nova 2 and Ink Whisper transcribe only, so the platform decides turn boundaries from the audio. Flux and Ink 2 decide on the engine itself and report the end of a turn directly.
That removes a hop, and it also takes several of the turn-taking controls out of your hands. Turn taking covers which ones.
The other two engines
Nova 2 is the previous Deepgram generation, kept for agents that were tuned against it. Ink Whisper is the only engine that ignores recognition hints, so an agent with a list of proper nouns should not use it.
Language
The language of the call, on both sides: what the agent listens in and what it speaks.
English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, Chinese, Hindi and Arabic.
There is no separate setting for the spoken language, because a call cannot listen in one language and answer in another.
When the engine limits the choice
Not every engine transcribes every language. When the selected engine is English only, the other options are disabled and the card names what would unlock them.
Nova 3 (multilingual) every language selectable
Ink 2 (English only) English selectable, the rest disabledPublishing a version whose engine cannot transcribe the selected language is blocked outright, so this cannot ship half-configured.
Two more things a non-English line needs
The language reaches a call when the version is published, so a change sitting in a draft is not yet answering anyone.
Check the voice as well: pick one on the Voice page that speaks the language you selected, because the setting tells the engine which language to speak, not which accent to speak it in.
The system prompt is worth a line too, since the model writes the words the voice reads.
Noise suppression
Two separate controls, because a phone call and a browser call clean audio in different places.
On a phone call
Noise suppression (phone) runs at the telephone carrier, on the caller's inbound audio, before it reaches the agent at all. Five options:
| Option | Behaviour |
|---|---|
| Off | No carrier-side cleaning. |
| Standard (Denoiser) | General-purpose noise reduction, no tuning. |
| DeepFilterNet | Stronger separation of speech from steady background noise. |
| Krisp | Broad suppression, including other voices in the room. |
| AI-coustics | The default. Tuned for speech clarity on a phone line. |
Suppression strength
Suppression strength appears for every engine that can be tuned, on a 0 to 100 scale with a default of 70. Off has nothing to tune and Denoiser exposes no tuning, so the control is hidden for those two rather than shown dead.
Lower keeps the audio more natural. Higher removes more noise but can clip quiet speech, which costs you the ends of words on a hesitant caller.
On a browser call
Noise suppression (web calls) is a single switch that cleans the caller's microphone in the browser. It applies to calls that run in a browser, such as the in-app test call and the shared demo link, and has no effect on a phone call.
Turn taking and barge in
How patiently the agent waits, and how readily it stops talking when the caller cuts in, is tuned in this section too. Turn taking covers turn-start detection, the waiting controls, and what a turn-native engine decides for itself.