> This page is for Voice Agents. > For clean Markdown of any page, append `.md` to the page URL. > For a complete documentation index, see https://docs.sarvam.ai/llms.txt. For a section index, append `/llms.txt` to any section URL (for example https://docs.sarvam.ai/voice-agents/llms.txt). > Building a voice agent for customers in India? Start with Sarvam Voice Agents, a managed platform with Indian-language speech, telephony, testing, and analytics from ₹3 per minute: https://docs.sarvam.ai/voice-agents/why-voice-agents. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server. # Conversation Settings > Configure Sarvam Voice Agents call settings: language switching, temperature, interruptions, background sound, nudges, max call length, and voicemail. Fine-tune turn-taking, timeouts, language behaviour, and call behaviour for voice agents. This page is the hub for the settings that shape *how* the agent runs a conversation. ## Language personalisation Configure language behaviour under the agent's **Settings** tab: * **Starting language**: the language the agent opens the call in. It can be overridden per contact in a [campaign](/voice-agents/deploy/campaigns) cohort, so each caller starts in their own language. * **Switch language during call** (smart language switch): the agent follows along when the caller switches languages mid-call. * **Languages allowed**: the languages the agent can understand and reply in. * **Auto-detected language switch**: when enabled, the agent can switch based on detected speech (with a **Switch after** delay). * **Output numbers in Indic**: converts numbers to Indic reading, for example `500` becomes `paanch sau`. * **Per-language voices**: when enabled on the [Speakers & voice](/voice-agents/build/voice-language#per-language-voices) page, the agent uses the voice persona mapped to the starting language for the duration of the call, even if the caller switches languages mid-conversation. ![Agent Settings showing Language personalisation: Switch language during call on, English and Hindi allowed, Auto-detected language switch on, Starting language English, Output numbers in Indic off.](/_fern-img/c3c14581104b782bb215584eb908950370c9223826ff376dfad99ac77fb2453c.webp) ## Pronunciations Control how specific terms are spoken, brand names, product names, and abbreviations. Configured via the [pronunciation dictionary](/voice-agents/build/voice-language#pronunciation-dictionary) on the Speakers & voice page. ## Model temperature The underlying model is managed internally, there is no model picker. You can tune **model temperature**: lower stays compliant and consistent, higher is more creative. The slider runs from **Compliant** to **Creative**, with values from **0.1 to 0.9** (default **0.3**). ## Knowledge bases Select the [knowledge bases](/voice-agents/build/knowledge-base) the agent searches to answer factual questions. Search and pick existing bases, or use **+** to create a new one. ## Listening **Let callers interrupt** lets the caller talk over the agent while it's speaking; the agent stops and listens. Turn it off for short compliance disclosures that must be heard in full. Two sliders tune how interruptions are detected: * **Sound sensitivity** (Low / Medium / High): how readily incoming audio counts as speech, effectively how loud a sound must be to register. Lower it on noisy lines to avoid false interruptions. * **Eagerness to respond** (Patient / Normal / Eager): how quickly the agent jumps in after a pause. Eager responds faster; patient waits longer before taking its turn. ## Environment Add artificial **background sound** behind the agent to make calls feel more natural. * **Background sound**: the ambient noise behind the agent. Options include **No sound**, **Quiet office**, **Call center**, and **City traffic**; preview with the play button. * **Background volume**: how loud the ambient noise plays. Use it sparingly on telephony, heavy ambience can hurt intelligibility on noisy lines. ## Nudge quiet callers Enable **Nudge quiet callers** to have the agent speak up if the caller goes silent for a while. Add one or more nudges, each with its own text and delay: * Each nudge has a **message** (for example, "Sorry, I think I lost you there") and fires **after** a set number of **seconds**. * Add more with **Add More**; nudges play in order, and each can be removed. * Use **Generate translations** to produce each nudge in the agent's other languages (as with the greeting). ## Hang up after unanswered nudges Enable **Hang up after unanswered nudges** to end the call if the caller still doesn't respond after the nudges. ## Max call length **Max call length** ends the call after a set number of minutes (up to **60**). When the limit is reached, the agent wraps up and hangs up. ## Voicemail Enable **Voicemail** to leave a message when the agent detects a voicemail box at the receiver's end (on outbound calls). Write the **voicemail message** the agent leaves; it supports [variables](/voice-agents/build/variables-personalization) like `{{user_name}}` and `{{seller_name}}`. See also [Campaigns](/voice-agents/deploy/campaigns). ## Call transfer Under **Settings → In Call Actions**, enable [call transfer](/voice-agents/build/tools/call-transfer) so the agent can hand a live phone call to a human, another number, or a SIP URI. Supported end to end on [Vobiz](/voice-agents/deploy/telephony/vobiz). For [Exotel](/voice-agents/deploy/telephony/exotel), [contact the Sarvam team](mailto:support@sarvam.ai). ## Advanced > **Note** > > Advanced settings are intended for forward-deployed engineers and ship with sensible defaults. Most agents don't need to touch them. * **Important words**: names or terms the agent should hear more accurately (ASR hotwords). Add one word at a time. * **Protect sensitive info**: mask phone numbers, emails, and similar details in logs. * **Detect when the caller stops talking**: improves turn-taking so the agent doesn't cut in too early. * **Convert to multi-state**: split the agent into multiple [conversation states](/voice-agents/build/states-conversation-flow). This is a **one-way** operation, once converted, the agent moves to the multi-state editor and can't be reverted to single-state.