Voice and language

A voice conversation is three steps in a loop: the caller's speech is transcribed, the model decides what to say, and the reply is spoken. This page covers the two ends.

Voices

Voices come from the speech providers Kilevel integrates, including ElevenLabs, PlayHT and Fish Audio. Fish Audio is offered as an EU/GDPR-oriented option for workspaces that need one. Custom voices can be added to a workspace as well.

Text is formatted for speech in the flow's language before it is synthesised, so numbers, dates and abbreviations are read out the way a native speaker would say them.

Languages

A flow has one primary language, which sets transcription, voice and the language the model answers in. More than 30 languages are supported.

Pronunciation

Brand names, product codes and unusual surnames are where synthetic voices stumble. Pronunciation dictionaries tell the voice how to say specific words, so the fix lives in one place instead of in phonetic spellings scattered through your prompts.

Interruptions and silence

Conversation feels natural or robotic mostly because of turn-taking, and these settings control it:

• Barge-in — whether, and how readily, the assistant stops speaking when the caller starts. A plain-text first message is spoken through by default.
• Speaking when unaddressed — whether the assistant speaks first when the conversation enters a new state, set per state.
• Waiting on tools — when a tool takes a while, the assistant can acknowledge the caller instead of going silent, and play a short sound cue while the action runs.