Skip to main content
Choose the speech and language options available in Klatt, then verify them with a workflow test.

​
Prerequisites

  • Access to the Klatt Console.
  • Organization access to Model Configurations.
  • A workflow that is ready for a short test call.

​
Choose a Klatt-managed model

  1. Open Models in the Console.
  2. Open the dropdown under Klatt-managed models and choose:
    • Gemma 4 31B — Recommended: faster replies with shorter pauses, suited to straightforward conversations that need less reasoning.
    • GLM 5.3 Flash: deeper reasoning for complex requests, with longer response times.
  3. Select Save model. Usage comes from your organization’s prepaid credits.
  4. Verify the choice with a short workflow test call.
Your saved choice remains selected until you change it. The recommendation does not replace an existing selection.

​
Speech recognition

The Gemma and GLM managed voice bundles use Deepgram Nova-3 for speech recognition with automatic multilingual transcription. Callers can speak or switch between English, Spanish, German, French, Italian, and Portuguese without changing the workflow language. Workflow language controls the default agent instructions and disclosures; Console language controls the interface. Noise suppression runs before recognition, and the call runtime handles interruptions and turn timing. Test names, numbers, pauses, and typical background noise for your use case; noise suppression does not isolate a speaker from every other voice. Managed recognition credentials are maintained by your platform administrator and are never included in a workflow. The agent does not play automatic acknowledgements over caller speech. It continues with contextual replies after the caller’s turn. Status audio for a slow read-only tool is available only while the caller is quiet.

​
Choose a model for one workflow

Open the workflow’s settings and enable Override for this workflow under Model Overrides. Use the same model dropdown and Save model action to choose Gemma or GLM for that workflow. To return to the organization choice, turn off the override and select Use Shared Configuration.

​
Hosted inference

Klatt is a hosted voice-AI product; customer self-hosted installations are not offered. Choose from the models available to your organization in the Console. Your platform administrator manages the provider credentials for Klatt-managed models. Do not put provider keys in workflow prompts or customer configuration.

​
Speaking speed and breath sounds

Open Models and use the Voice delivery card below the managed model card. These settings apply to your organization’s managed Fish voices on new calls, including workflows with their own model selection.
  1. Move Speaking speed from 0.7× to 1.3× in 0.1 steps. The filled control shows the current value; 1.0× is normal and 1.1× is the default.
  2. Turn Natural breaths on or off. It is off by default.
  3. Select Save voice settings, then test a workflow. Saving this card keeps the saved model selection. Saving the model card keeps saved voice settings.
The slider supports dragging and arrow keys. Its settings are available when the saved managed model uses Fish; other voices retain their own delivery. Speaking speed changes delivery, rather than the time the agent takes to start answering. The result varies with the voice, language, and sentence. The current gateway uses OpenRouter’s documented speed parameter, whose support depends on the selected provider. Natural breaths is an optional setting for Fish S2.1 Pro. When enabled, Klatt adds delivery guidance allowing occasional [inhale] or [exhale] sounds. Most replies should have no cue. Klatt keeps these markers out of live and saved assistant transcripts and conversation history; callers’ words and tool results are not rewritten. You do not need to add these markers to workflow prompts. Keep prompts focused on short, natural sentences and the conversation’s purpose. The delivery guidance excludes breaths from disclosures, safety instructions, numbers, dates, prices, identifiers, codes, tool output, and backchannels. Test the audio before enabling breaths: accepting a marker does not guarantee a natural sound.

​
Constraints and failure states

Select an available model and save your choice. Models whose provider credentials are not ready cannot be selected; contact your platform administrator if a model is unavailable. After changing an organization model or workflow override, run a short workflow test before using it for calls.

​
Protect your data

Never paste a provider key into documentation, browser code, workflow prompts, or a support message. Contact your platform administrator if a managed provider is unavailable or its credentials need to be replaced.