Skip to main content
Why VAW Native

Curated AI Models, Voices and Languages

The AI models, voices, voice models, languages and conversation controls you can choose from on VAW Native, and why the model list is curated for voice.

Written By voiceaiwrapper

Last updated 12 days ago

On VAW Native, the AI models, voices and speech recognition run on VoiceAIWrapper's accounts, so you don't need keys from any model or voice provider. You pick from a list chosen for live calls. Each section links to the article for that setting.

Applies to: agents on a VAW Native Pod (an engine or provider connection). On Vapi, Retell AI and ElevenLabs Pods, you change an agent's model, voice and language in the provider's own dashboard. See Provider Feature Comparison.

You'll find all of these on the agent's Agent Settings tab, in the Voice, Language, AI model and Conversation rows.

AI models

The AI model list has 32 models from three providers:

ProviderModelsExamples
OpenAI20GPT-4o mini, GPT-4.1, GPT-5 mini, GPT-5.6 Luna (the default for new agents)
Anthropic7Claude Haiku 4.5, Claude Sonnet 4.6, Claude Opus 5.5
Google5Gemini 3.1 Flash-Lite, Gemini 3.8 Flash, Gemini 4 Argon

Select AI model lets you search, filter by provider and compare the credits/min next to each model: an estimate of what the model alone uses in a typical minute, charged at cost with no markup. A call's total also includes the voice, speech recognition and the platform fee.

Agent page β†’ Agent Settings β†’ AI model β†’ Select AI model, with the provider filter set to Anthropic: Search models…, the provider filter (All, OpenAI, Anthropic, Google, each with a count), the Anthropic models each showing {cost} credits/min, and the estimate note at the bottom of the list.

Post-call analysis has its own model setting on the Analysis tab, with the same list plus one analysis-only model, GPT-5.6 Sol. See Choose an AI Model.

Why the list is curated

  • Speed on live calls. Most models on the list are reasoning models. On calls, they always run at a fixed low reasoning setting, because more reasoning adds seconds before the agent's first word. Their sheet says: "This model is set up for speed, so the caller is not kept waiting."
  • Every model can be costed. Every model on the list, including the analysis models, has a rate on file, so each call's What ran tab can break down what the model cost.
  • Replaced models are left out. For example, Gemini versions that Google has deprecated in favour of newer ones aren't offered.

Voices

The voice library has 42 ElevenLabs voices: 18 general voices, plus 24 voices made for Korean, Russian, Indonesian, Italian, Turkish, Japanese, Dutch, Polish, Danish, Hungarian, Portuguese (Brazil) and Portuguese (Portugal).

  • Find a voice with Search voices…, the language button and Filters (Speaks, Gender, Accent, Recommended only).
  • Preview any voice with its play button before you choose.
  • Compare cost: each tile shows "from X credits/min".
  • Choose a voice model (six ElevenLabs models), sorted by Recommended, Fastest, Most natural or Lowest cost.
  • Tune the delivery with Voice speed, Stability and Similarity, and add an optional Background sound.
Agent page β†’ Agent Settings β†’ Voice β†’ Select a voice, with the Filters menu open (Speaks, Gender, Accent, Recommended only) and one filter chip active, the {shown} of {total} voices counter, and a voice tile's play button in its playing state.

OpenAI voices are no longer offered for new choices. See Choose a Voice for Your Agent.

Languages

The Language setting offers Auto-detect plus 55 languages and regional variants, such as English (UK) and Portuguese (Brazil).

  • Auto-detect covers English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch.
  • Twenty languages, including Hebrew and Thai, can only be spoken by the Eleven v4 Turbo, Eleven v3 Conversational and Eleven v3 voice models.
  • The language controls how the agent understands callers. The agent replies in the language of its prompt and greeting, or the one its prompt tells it to use.

See Choose the Agent's Language.

Conversation controls

The Conversation sheet sets how the agent takes turns and understands callers:

ControlWhat it changes
Response patienceHow long the agent waits after the caller stops talking: Fast, Balanced or Patient.
Allow interruptions and Interruption sensitivityWhether callers can talk over the agent, and how easily.
Background noise filterReduces background noise and other voices on the caller's side.
Boosted wordsUp to 25 names and terms to recognise accurately. Ignored on Auto-detect.

See Fine-Tune Conversation Behaviour.