> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.rydhm.ai/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.rydhm.ai/_mcp/server. # Voice engines The **voice engine** decides how an expert turns a caller's speech into its own. On the expert's **Voice** tab you pick a **Style**, and each style runs one engine with a tested set of models: | Style | Engine | What it's for | | ---------------- | -------------------------------------------------------- | ------------------------------------------------------------------------------- | | **Fastest** | **Realtime (Google)**: Gemini Live | The shortest pause before it answers. Google's voices. | | **Balanced** | **Standard**: transcriber → language model → voice model | The default. Steady with bookings and other actions, in calls and meetings. | | **Most natural** | **Realtime (OpenAI)**: GPT-Live | Listens while it talks, like a person. One-on-one calls; meetings use Balanced. | Each style card shows its measured pause before a reply. To choose the engine and its models yourself, open **Advanced** → **Voice engine**. If you change a model by hand, the style reads **Custom**; pick a style to go back to a tested setup. ![The Voice tab: three styles, each with its estimated pause before a reply, and Advanced open on the three voice engines](/_fern-files/_img/1bc4484c773585fa6dfc35d381063b089f02a6ea5949dd64a3c676b61a57f1b5.webp)![The Voice tab: three styles, each with its estimated pause before a reply, and Advanced open on the three voice engines](/_fern-files/_img/4934835496b250f68bfb158ccedd4e0f4ddc0f48db6b21dc9b93a6c5c6feb124.webp) | | **Standard** | **Realtime (Google)** | **Realtime (OpenAI)** | | ---------------------------------- | ------------------------------------------------------------------ | --------------------------------- | -------------------------------------------------- | | How it works | Transcriber → language model → voice, each picked separately | One Google model hears and speaks | OpenAI's full-duplex model listens while it speaks | | Voices | Cartesia and Deepgram | Google's 30 voices | OpenAI's 31 voices | | Meetings | Yes | Yes | No — meetings run Standard | | Voice controls | Applied exactly | Followed closely, not exactly | Followed closely, not exactly | | Backup voice if the provider fails | Yes | No | In meetings only | | Deep thinker | Yes | Yes | No (it has its own thinking model) | | Editor's estimate of the pause | The sum of the three parts you pick; about 1.5 s with the defaults | About 0.7 s | About 1 s | **Standard** (the **Balanced** style) is the default. Switching engine keeps the other engines' settings on the expert, so switching back restores them exactly. An engine has its own voices, so switching also moves the expert to the closest voice (same gender, language and accent), and the editor tells you which. ## Standard Choose Standard when you want control over every stage, the widest choice of voices, or the exact voice controls. In the app it is the engine behind **Balanced**; the vendor names below appear only under **Advanced**. You pick three things: * **Transcriber**: speech-to-text. *Deepgram Flux — English* is the default and takes turns fastest. The others are Flux multilingual (for callers who mix languages in one sentence), Deepgram Nova-3 in English or multilingual, and Gemini 3.5 Transcribe Live, which detects any language. * **Language model**: the model that writes each reply. The list groups models by provider (OpenAI, Anthropic, Google). Each row shows its typical and worst measured time to first word, and a cost tier: **Lowest cost**, **Mid cost** or **Highest cost**. New experts start on **GPT-5.6 Luna**. If you want Claude on voice, the picker recommends **Claude Haiku 5.5**. * **Voice model**: **Cartesia Sonic-3.6** (the default), Sonic-3.5 or Sonic-3, or **Deepgram Aura-2 & Flux TTS**. The voice you choose decides which. > **Tip** > > Read the **worst** time, not just the typical one. Most models start in a similar band; their slowest turns differ far more, and the slowest turn is the one a caller remembers. The times are measured with a knowledge base in the prompt, the way experts really run. If the provider of a model you picked announces its retirement, the field says when, and which model takes over, for example *"Retires Jan 14, 2027, then runs as GPT-5.6 Luna."* The expert keeps working on the successor; nothing breaks on the day. ## Gemini Live **Realtime (Google)** in the editor, the engine behind **Fastest**. Choose it for quick, natural-sounding replies and for callers who switch language: it answers in the caller's language by itself once you allow that language. * One model (`gemini-3.8-live`) hears and speaks. There is no language model or voice model to pick, so those fields are hidden. * **Transcriber (meetings)** stays: the transcriber decides when a caller's turn has ended, and in meetings it decides when the expert is being addressed. Gemini only produces the voice. * Voices come from Google's set of 30. Each speaks whatever language the conversation is in. * **Speed**, **Tone** and **Pronunciations** become instructions to the model, so it follows them closely but not exactly. There is no backup voice. ## GPT-Live **Realtime (OpenAI)** in the editor, the engine behind **Most natural**. Choose it for the most natural one-on-one conversations. It listens while it speaks and decides for itself when to talk, pause or let the caller in. * **One-on-one only.** GPT-Live doesn't join Meet, Teams or Zoom calls. In a meeting the expert runs the Standard settings you set under **In meetings**: a **Meeting voice**, a transcriber, a language model and a voice model. * **Thinking model**: GPT-Live hands anything that needs reasoning or a tool to this model, and keeps talking while it works. | Thinking model | When to pick it | | ------------------------ | ------------------------------------------------- | | **GPT-6 Luna** (default) | Fast and cheapest. Right for lookups and booking. | | **GPT-5.6 Luna** | The previous fast model, at twice the price. | | **GPT-6.1 Terra** | More careful with multi-step tools. | | **GPT-6 Sol** | Strongest reasoning; about 20× Luna per answer. | * The voice stays the same for the whole conversation. Some voices are made for one language, such as Hindi, German, French, Japanese or Korean. * Voice previews take a few seconds. ## Hear it before you save At the bottom of the **Voice** section, **Hear the first message** plays the expert's greeting in the voice and language a caller would get, with your voice controls applied. Beside it, **Est. pause before reply** shows the editor's estimate for the whole setup. To see how fast your experts really answer, open [Analytics](/quality/analytics). ## Cost Rydhm prices a call by the minute, and **the minute price is the same whichever style or engine the expert uses**. See [Billing](/workspace/billing). The cost tiers on language models and the notes on thinking models compare what the models cost to run, to help you pick the cheapest one that does the job. If your workspace adds its own OpenAI or Google key, experts on that provider run on it. See [Provider keys](/workspace/provider-keys). ## Gotchas * **Keep the Standard settings sensible on every expert.** If Google or OpenAI can't be reached when a call starts, a Gemini Live or GPT-Live expert takes the call on Standard instead of staying silent. * **Languages and the transcriber move together.** See [Languages](/experts/languages). * **A voice from another provider switches the engine.** See [Voices](/experts/voices). ## Related * [Voices](/experts/voices) * [Languages](/experts/languages) * [Guardrails and call experience](/experts/guardrails) * [Provider keys](/workspace/provider-keys)