> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ouraicalling.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Models & voices

> The model catalog, the voice library with previews, and fallback chains

## The model catalog

Assistants don't hardcode providers — they pick from a **curated model catalog** covering four types:

| Type           | Providers (examples)                        |
| -------------- | ------------------------------------------- |
| LLM            | OpenAI, Anthropic, Google, Groq, Azure      |
| Speech-to-text | Deepgram, OpenAI, Google, Azure             |
| Text-to-speech | Cartesia, ElevenLabs, OpenAI, Google, Azure |
| Realtime       | OpenAI (gpt-realtime), Google (Gemini Live) |

Only the **platform admin** controls availability. They enable models globally,
can grant or block a model for one user, and define which engine modes each model
supports. Tenant admins cannot override the catalog.

With the Fallbacks & Guardrails add-on, the assistant editor shows provider and model selectors.
Pipeline, realtime, and half-cascade keep independent choices. Choosing the
default stores `null`, so later admin changes take effect automatically:

1. Assistant selection (Fallbacks & Guardrails only)
2. Personal default set by the platform admin
3. Platform default
4. Safe runtime fallback

Without Fallbacks & Guardrails the editor remains automatic. Personal admin defaults still apply.
Unavailable legacy values remain visible so they can be reset, but cannot be
selected again.

<Tip>
  Look for the **Recommended** badge — those are the models we test most and tune defaults for. **Low latency** flags models suited for snappy phone conversations.
</Tip>

## The voice library

The voice picker loads the complete paginated catalog from every enabled TTS provider, filterable by language and gender, **with instant audio previews** — click play to hear a sample before committing.

ElevenLabs voices with custom rates or credit multipliers are excluded, so every ElevenLabs voice shown in the picker uses the provider's standard voice rate.

* **Voice per assistant** — the default speaking voice.
* **Voice override per flow agent** — individual [flow](/flow-builder/nodes) agents can speak with a different voice (e.g. a distinct voice for the "sales specialist" agent).
* **Voice per language** — with [automatic language switching](/assistants/languages), you map a voice per language so the assistant switches voice together with language.
* **Cloned voices** — with the **Clone your own voice** add-on (or Include free on the plan), you can add custom clones and use them like any library voice. A clone is private to the account that created it: other members of the same workspace cannot see or select it. Capacity is limited by `max_cloned_voices`.

## Speaking style

Pipeline and half-cascade assistants use the same provider-aware voice controls. The editor only sends settings supported by the resolved TTS model:

* **Cartesia Sonic 3 / 3.5** — speaking rate, output volume, and optional dynamic emotion cues.
* **Eleven v3** — stability and native audio-tag expressiveness. Unsupported speed, similarity, style, and speaker-boost controls are not sent.
* **Other supported ElevenLabs models** — speaking rate, stability, similarity, style exaggeration, and speaker boost.
* **OpenAI and Gemini TTS** — free-text speaking-style instructions when the selected model supports them.
* **Pronunciation dictionary** — per-assistant key-value mapping, e.g. `"API" → "A P I"` or your company name spelled phonetically. Applied to all spoken output.
* **Markdown & emoji filters** — enabled by default so LLM formatting artifacts are never read aloud.

Changing any effective voice control also changes the phrase-cache key, so cached greetings and filler phrases cannot reuse audio produced with older settings.

## Fallback chains (Fallbacks & Guardrails)

For production reliability you can define a **fallback chain per stage**: if the primary provider degrades mid-call, the engine switches to the next entry — the call continues seamlessly.

* **STT** — e.g. `Deepgram nova-3 → AssemblyAI universal-streaming`
* **LLM** — e.g. `GPT → Groq`
* **TTS** — e.g. `Cartesia sonic → ElevenLabs` (each entry may specify its own voice)

Provider health changes are logged as call events, so you can see afterwards whether a fallback fired. Fallbacks are a plan-gated feature.

<Note>
  Even without configured fallbacks, a broken configuration never kills a call: the engine falls back to a safe default stack and records a `config_error` event.
</Note>