The call voice
In Configuration → Assistant, Voice tab, you choose how the assistant sounds when someone in your organization calls it from the panel (the wave button next to the text box). There are two parts:
| Part | What it does |
|---|---|
| Hearing (transcription) | Turns what the person says into text. |
| Voice (synthesis) | Reads the assistant’s answers aloud. |
For each one you can keep Local (automatic, no cost) or choose a provider. Click Save voice when you finish; the next call uses it.
Local, at no cost
The default. Transcription and voice run on the NappAI server and cost nothing. They sound somewhat more artificial than the paid providers.
A paid provider
You can use the same providers as a voice agent: Deepgram, Cartesia, ElevenLabs, Inworld, Azure, OpenAI, Gemini… and also the local engines that are not the default, such as Kokoro or Supertonic.
- Choose the provider.
- Choose the credential. Your organization’s credentials of the type that provider needs are listed. If there are none, create one in Credentials and come back.
- Optionally change the model, the voice or the speed. Whatever you leave empty uses the provider’s default, the same as in a voice agent.
The call’s language is the caller’s platform language. With some providers the voice carries its language (for example, Deepgram voices ending in -es are Spanish).
What it spends is recorded against that credential in Usage and counts towards its spending limits: transcription in seconds of audio and voice in characters read.
If the chosen voice fails
If the credential no longer exists or is not valid, the provider does not answer or the credential has reached its spending limit, the call carries on with the local voice. The call panel says so: «The configured voice isn’t available; using the local one», with the reason. The same happens if the provider fails mid-call.