Real-Time Voice

Speak naturally, get spoken replies — under-second latency for support, sales, and in-app agents.

AresGen connects Cartesia Sonic streaming TTS, Deepgram Nova-3 speech-to-text, and ElevenLabs cloned voices into one live voice loop. No buffering. No robotic pause. Just conversation.

Start free Read docs

Under-second response · 60+ locales · Streaming transcripts

Streaming transcription via Deepgram Nova-3 — partial results while speaking
First-phoneme TTS under 200ms with Cartesia Sonic
Interruption-safe — user barge-in stops the bot mid-sentence
Use any voice in your AresGen workspace — no separate vendor accounts needed

00:42EN-US · Sarah · Calm02:18

End-to-End Latency

<800ms

From mic to spoken reply, streaming

Locales

60+

Across speech-to-text and TTS engines

Audio Engines

Hot-swap Cartesia / Deepgram / ElevenLabs

Capabilities

Built for production teams.

Streaming STT with partial results

Deepgram Nova-3 delivers rolling transcription during speech — no waiting for end-of-utterance. Partial results feed the LLM mid-sentence for lower total response time.

First-phoneme TTS under 200ms

Cartesia Sonic begins streaming audio before the full text is generated. ElevenLabs handles cloned voices with equivalent streaming latency. Both engines integrate without code changes.

Interruption-safe barge-in

When the user speaks over the agent, the audio stream is halted immediately. Conversation context is preserved — the agent picks up from the right place without losing turn memory.

Conversation memory across turns

Turn-by-turn transcript is stored in the session context. The agent carries brand voice profile, user preferences, and prior turns into every reply without manual state management.

Session transcript export as JSON or CSV

Every voice session produces a full rolling transcript via Deepgram Nova-3 STT. Export the completed transcript as JSON or CSV for CRM import, compliance archiving, or QA analysis — no manual transcription required.

Function calls and tool use during voice turns

The AI Chat reasoning layer runs behind every voice session — invoke tools, query external APIs, or run structured function calls mid-conversation. Results are spoken back in the same turn without interrupting the voice loop.

Live agent assist

Give your support team a real-time voice copilot.

Agents on live calls get live transcription, suggested replies, and instant knowledge base lookups — all surfaced in a side panel while the call is happening. No manual notes. No hold time while searching docs.

Live transcription via Deepgram Nova-3 during the call

Suggested replies generated from brand knowledge base

Auto-summarise call after hang-up — no manual notes

Escalation detection — flag sentiment shifts in real time

Works alongside your existing call infrastructure without replacement

Explore AI Chat for the reasoning layer

Voice — Aria · neutral · EU-EN

0:14 / 0:42

transcript — "Welcome back. Today we're walking through three patterns teams use to ship AI agents that don't embarrass them in front of customers…"

In-app voice agent

Embed a hands-free voice widget in your SaaS.

Drop a voice widget into any web or mobile app. Users ask questions aloud and hear answers spoken back — no clicking, no typing. Ideal for accessibility, field use cases, and hands-free onboarding flows.

Embed via lightweight SDK — under 10 KB on the wire

Multi-language support across 60+ locales

Custom wake-word or push-to-talk activation

ElevenLabs cloned voice for consistent brand personality

Session transcript exported automatically after each conversation

See support deployment patterns

Voice — Aria · neutral · EU-EN

0:14 / 0:42

transcript — "Welcome back. Today we're walking through three patterns teams use to ship AI agents that don't embarrass them in front of customers…"

Use cases

See it in action.

Deploy a 24/7 customer-service voice bot that sounds human.

Prompt

Answer inbound calls about shipping delays in English, Spanish, and French. Escalate to a human agent if sentiment drops below neutral.

Sample output

[Voice bot activated] "Your order is in transit and will arrive by Thursday. Would you like a tracking link sent to your email?" — sentiment score: positive. Escalation: not triggered. Call duration: 42 seconds.

Qualify inbound leads with a spoken discovery call before handoff.

Prompt

Run a 5-question discovery call for enterprise prospects. Capture company size, use case, and budget range. Route high-fit leads to AE calendar.

Sample output

[Lead profile captured] 200+ seats, use case: internal helpdesk, budget: above threshold. Calendar invite sent to account executive. Call summary written to CRM.

Let users navigate complex SaaS interfaces entirely by voice.

Prompt

User says: "Go to the last invoice and download it as a PDF." App has no keyboard shortcut for this action.

Sample output

[Voice agent parsed intent → navigated to invoices → opened invoice #1042 → triggered PDF export] "Done — your invoice is downloading now." Action completed in 3 seconds.

Run live multilingual interview sessions with real-time interpretation.

Prompt

Interview a Portuguese-speaking candidate in real time. Translate questions to Portuguese, capture answers, translate replies back to English for the hiring panel.

Sample output

[Bidirectional voice interpretation active] English question spoken → Portuguese TTS for candidate → candidate replies → Deepgram transcription → English translation read back to panel. Zero-lag bilingual session.

Pairs well with.

AI Chat

The conversational model behind every voice turn — system prompts, tools, and conversation memory feed the realtime loop.

Learn more

Voiceover

When you need pre-recorded narration instead of streaming, Voiceover is the batch sibling — same voices, deeper control.

Learn more

AI Writer

Draft system prompts, call scripts, and reply templates before they ever reach the voice loop.

Learn more

Solutions for Support

Support persona — agent assist, live voice bots, and ticket deflection built on AresGen tools.

Learn more

Frequently asked

What is the real end-to-end latency?

Under 800ms in typical conditions — measured from final user phoneme to first spoken reply phoneme. Deepgram Nova-3 delivers partial transcripts during speech, so the model can start reasoning before the utterance ends. Cartesia Sonic begins streaming audio before the full response text is ready. Network round-trip from a Hetzner-hosted deployment to EU clients is typically under 50ms.

Can I switch voice engines mid-session without downtime?

Yes. Cartesia Sonic, Deepgram Nova-3, and ElevenLabs are configured as hot-swappable providers in AresGen. You can change the TTS or STT engine from the workspace settings panel without restarting the session or modifying code. Switching engines resets audio state but preserves conversation context.

How does barge-in work — can the user interrupt the agent?

Yes. When the voice activity detector registers speech from the user, the current audio stream is stopped immediately and the agent begins processing the new input. Context from the interrupted turn is retained so the agent does not lose track of what it was saying. Barge-in sensitivity is adjustable per deployment.

Is conversation history stored and is it private?

Conversation transcripts are stored in your AresGen workspace only. They are encrypted at rest, scoped to your account, and are never used for model training. You can configure automatic transcript deletion after a set retention window. For GDPR compliance, all storage remains in your chosen Hetzner region.

Launch a real-time voice agent in minutes — no telephony expertise required.

Start free. Connect Cartesia, Deepgram, and ElevenLabs in a single workspace. Your voice loop is live before your next meeting.

Start free View pricing