Skip to main content

Models and voices

Open an agent and choose Model & Voice to choose what powers it. An agent has two separate engines:

  • A chat model for text conversations.
  • A voice provider for phone calls and browser voice calls.

Changing one doesn't change the other. The panel next to the settings shows how a call flows with your choices and what each one costs. For prices, see voisx.ai/pricing.

Chat model​

The chat model answers text conversations: the website widget chat, WhatsApp, web forms and workflows.

SettingWhat it does
ProviderThe model provider, for example OpenAI (GPT) or Google Gemini.
ModelThe model to use. Provider default uses the provider's default model.

The lists show the providers and models available on the platform. Switching provider clears the model, so an old model never carries over.

Voice​

The voice provider runs phone calls and browser voice calls. There are two kinds:

KindProvidersHow it works
Realtime speech-to-speechOpenAI Realtime, Google Gemini LiveOne model hears the caller and speaks the reply in a single step. This gives the most natural turn-taking. OpenAI's gpt-live-1 model goes further: it listens and speaks at the same time (full duplex). See GPT-Live voice.
Speech-to-text → chat → text-to-speechElevenLabs, SarvamThe caller's speech is transcribed, the agent's chat model writes the reply, and the voice provider speaks it.
SettingApplies toWhat it does
ProviderAllThe voice provider.
Voice modelOpenAI Realtime, Gemini Live, ElevenLabs, SarvamFor realtime providers, the realtime model, for example a full or a "mini" model. For OpenAI, the list can also include gpt-live-1, the GPT-Live model. See GPT-Live voice. For ElevenLabs and Sarvam, the text-to-speech model. See ElevenLabs models and Sarvam. Default uses the provider's default.
Transcription modelOpenAI Realtime, ElevenLabs, SarvamFor OpenAI Realtime, transcribes the caller's speech for transcripts and analytics. The default is whisper-1 on phone calls and gpt-4o-mini-transcribe in the browser. You can also choose gpt-4o-transcribe. For ElevenLabs and Sarvam, it turns the caller's speech into the text the chat model answers: OpenAI (default) or Sarvam Saaras. See Sarvam transcription.
Voice personaAllThe voice callers hear.

If a voice model you chose is later removed from the platform, the agent page shows a warning. Live calls keep using it until you choose another model.

Voice personas​

Each provider has its own voices. When you switch provider, the agent switches to that provider's recommended voice.

ProviderVoicesRecommended
OpenAI Realtimealloy, ash, ballad, coral, echo, sage, shimmer, verse, marin, cedarmarin, cedar
OpenAI, model gpt-live-1The ten OpenAI Realtime voices, plus 21 regional and language voices, including Nira (Hindi). See Voices.marin, Nira
Google Gemini Live30 voices, including Puck, Charon, Kore, Zephyr, Aoede and LedaPuck, Charon, Kore
ElevenLabsThe full ElevenLabs voice library, searchable. See ElevenLabs voice library.Featured voices, shown first
SarvamThe Sarvam speakers for the chosen model and language. See Sarvam voices.Shubh

A voice the provider no longer offers shows as "unavailable". Pick another one.

ElevenLabs models​

An ElevenLabs agent can choose which ElevenLabs text-to-speech model speaks its replies. The Voice model list shows the models that are available on the platform.

ModelLanguagesPrice tier
eleven_v494Premium
eleven_v4_turbo94Standard
eleven_v374Premium
eleven_v3_conversational74Standard
eleven_multilingual_v229Premium
eleven_flash_v2_532Standard
eleven_turbo_v2_532Standard
eleven_flash_v2English onlyStandard
eleven_turbo_v2English onlyStandard
  • New agents start on eleven_v4_turbo.
  • Existing agents that never chose a model keep using eleven_flash_v2_5. Nothing changes for them until you pick a model.
  • The model must speak the agent's main language and every language the agent may switch to. If it doesn't, the editor warns you and saving fails with a message such as "eleven_flash_v2 does not speak Arabic." Choose a model with more languages, or change the agent's language settings.

ElevenLabs calls are billed per minute in one of two tiers:

TierModels
ElevenLabs Standardeleven_v4_turbo, eleven_v3_conversational, eleven_flash_v2_5, eleven_turbo_v2_5, eleven_flash_v2, eleven_turbo_v2
ElevenLabs Premiumeleven_v4, eleven_v3, eleven_multilingual_v2

Premium models cost more per minute than standard models. Your rates are on the Usage rates tab of the billing page. See Usage and billing and voisx.ai/pricing.

tip

eleven_v4_turbo is a good default: it covers 94 languages at the standard rate. Choose a premium model when voice quality matters more than cost.

ElevenLabs voice library​

With ElevenLabs you can choose any voice from the ElevenLabs voice library, which has thousands of voices. In Model & Voice, open the voice picker to search it:

FilterWhat it does
SearchMatches the voice's name, description and use case.
LanguageVoices recorded in that language.
AccentVoices with that accent, for example a Gulf Arabic or an Indian English accent. The accents listed depend on the language you chose.
GenderMale, female or neutral voices.
AgeYoung, middle-aged or old voices.
  • Press play on a voice to hear a sample before you choose it. Playing a sample is free.
  • Featured voices appear first, before you search or filter. They are voices the VoisX team recommends.
  • Some voices are not offered on VoisX. They don't appear in the picker and can't be chosen.

The agent page shows the voice the agent uses now, even if it's not on the current page of results.

note

Gemini Live can't be used on Plivo phone numbers yet. For an agent on a Plivo number, choose OpenAI Realtime, ElevenLabs or Sarvam.

Sarvam​

Sarvam AI (India) is a speech provider built for Indian languages. With Provider set to Sarvam, Sarvam Saaras transcribes the caller, the agent's chat model writes the reply, and Sarvam Bulbul speaks it. It works in the website widget, the Playground and on phone calls (Twilio, Plivo and SIP trunks). Calls always run through the VoisX relay; the browser never connects to Sarvam.

Sarvam must be enabled

Sarvam is off until the VoisX platform team enables it, and it's available only on plans that include Sarvam voice. If Sarvam isn't in the Provider list, or saving the agent fails with "feature not in plan", contact us to turn it on.

Sarvam models​

ModelName in the listLanguages
bulbul:v4-flashBulbul v4 FlashHindi, Bengali, Tamil, Telugu, Gujarati, Kannada, Marathi, Punjabi and English (India). Not Malayalam or Odia.
bulbul:v3Bulbul v3Hindi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Odia and English (India)
  • New Sarvam agents start on bulbul:v4-flash.
  • If the agent's main language is Malayalam or Odia and you didn't choose a model, the agent uses bulbul:v3 automatically.
  • If you choose bulbul:v4-flash for an agent whose main language, or a language it may switch to, is Malayalam or Odia, saving fails with a message such as "bulbul:v4-flash does not speak Malayalam." Choose bulbul:v3, or change the agent's language settings.

Sarvam voices​

The voice picker lists the Sarvam speakers for the model you chose, in the agent's main language.

ModelVoicesDefault
bulbul:v337 speakers, including Shubh, Aditya, Ritu, Priya, Neha, Rahul, Kavya and Rohan.Shubh
bulbul:v4-flashVoice personas named by speaker, language and style, for example shubh_enhi_ads (Shubh, Hindi-English).shubh_enhi_ads

Each model has its own speakers. When you switch model, choose a voice from the new model's list.

Sarvam transcription​

Sarvam Saaras can transcribe the caller for any speech-to-text agent, on Sarvam or ElevenLabs. Choose it in Transcription model:

OptionModel
OpenAI (default)The platform's OpenAI transcription.
Sarvam Saaras v4sarvam:saaras:v4, the recommended Sarvam model.
Sarvam Saaras v3 Realtimesarvam:saaras:v3-realtime

Sarvam Saaras is built for Indian languages, including callers who mix languages, such as Hindi and English. Like the Sarvam voice, it can be chosen only when Sarvam is enabled for your organization and included in your plan.

Sarvam voices and transcription are billed by part, not per minute. See Usage and billing.

Language​

Each agent has one Language section, in Overview. It has three settings:

SettingWhat it does
Main languageThe language the agent greets in and starts every conversation in. The list shows only the languages the agent's voice speaks.
Stay in <main language>The agent always replies in its main language, whatever language the caller speaks.
Follow the callerThe agent starts in its main language, then replies in the caller's language once they keep speaking it.
Languages it may switch toShown for Follow the caller. Pick the languages the agent may switch to. Leave them all unpicked to allow any language the voice speaks. The main language is always allowed.

With Follow the caller, the agent switches when:

  • the caller keeps speaking another language for more than one turn, or
  • the caller asks for another language, for example "can you speak Kannada?".

A single word, a name or a greeting in another language doesn't switch the language. When the caller goes back to the main language, the agent goes back too.

Which languages a voice speaks​

VoiceLanguages
OpenAI Realtime, GPT-Live, Gemini LiveA broad list: English, the main Indian languages (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia and Urdu), Arabic, and more than 30 others.
ElevenLabsThe languages of the chosen ElevenLabs model.
SarvamThe languages of the chosen Sarvam model. bulbul:v4-flash doesn't speak Malayalam or Odia.

The main language and every language the agent may switch to must be ones the voice speaks. If you change the voice in Model & Voice to one that doesn't speak them, the editor warns you and you can't save until you change the voice or the language settings. Model & Voice also shows what the agent speaks, for example "Speaks: Hindi only · change on Overview". You change the language only in Overview.

Language rules in the system prompt​

The Language setting overrides any language rule in the system prompt, such as "Always reply in Hindi". If the prompt looks like it has one, the editor shows a hint. Remove the rule from the prompt, or make the setting match it.

Agents created before the Language section​

  • An agent with no language set now uses English as its main language, with Follow the caller. It behaves as before: it answers in the caller's language.
  • An agent with a language set now uses Stay in <that language>. Choose Follow the caller if it should switch.

Try a different voice without saving​

In the Playground, the Test-session voice settings let you try another voice model, voice and language for one browser session without changing the agent. The test call runs through the VoisX relay, the same path as production calls. See Playground.

Tips for fast, natural calls​

  • Realtime models give the most natural turn-taking. "Mini" models cost less and are usually fast enough for simple tasks.
  • Keep the system prompt focused. Shorter prompts add less delay to each reply.
  • Ask for short replies in the prompt. One or two sentences per turn works well on a call.
  • Tune interruptions and pauses in Call settings.
  • For prompt rules and tool settings that work well on calls, see Voice best practices.
Was this page helpful?