Speech-to-Speech AI Voice Agents
Human-like AI voice agents on native speech-to-speech models or STT → LLM → TTS pipelines, chosen call flow by call flow and running in production.
02 Feature
Voice AI agents in French, Dutch, Spanish, Portuguese and English.
Each CallShift agent speaks one of five languages, with voices, templates, messages and spoken dates localized, and French tuned the deepest. French and Spanish run at scale in production.
01 In short
Which languages do CallShift voice agents speak?
A CallShift agent speaks English, French, Dutch, Spanish or Portuguese (Brazilian). Each agent runs in one language, and the language sets more than a flag: the recognition language, the default voice, the idle and closing phrases, and the language of templates, messages, summaries and spoken dates.
French and Spanish run at scale in production today, and French is the language we tune the deepest.
- 5languages: English, French, Dutch, Spanish and Portuguese
- 27voices tagged for French, the deepest tuning of the five
- 100clubs on French and Spanish agents at OnAir Fitness
Updated 30 September 2026
02 Language by language
What each language gets.
| Language | Voices | What to know |
|---|---|---|
| English | 29 | Localized templates, messages and spoken dates. |
| French | 27 | Tuned the deepest: turn-taking and listening cues. |
| Dutch | 8 | Localized templates, messages and spoken dates. |
| Spanish | 6 | One Spanish, with no regional variant. On speech-to-speech models, accent and style are steered from the prompt. |
| Portuguese | 5 | Brazilian Portuguese. |
03 What is localized
More than a setting.
- The recognition language and the default voice, chosen with the language.
- The idle, silence and closing phrases of the agent.
- Templates, tool messages, SMS and email replies, summaries, and spoken dates and times.
- A pronunciation glossary per agent, for names and terms that must be said right.
- A choice of speech-to-text providers, including European ones such as Mistral and Gradium.
04 One language per agent
How to run several languages.
An agent speaks one language for the whole call: it does not detect the language or switch in the middle. To serve several languages, run one agent per language.
For dictated emails and numbers, the agent reads them back before confirming, and we use a pipeline where that precision matters most.
05 In production
Where each language runs.
OnAir Fitness runs French and Spanish agents across 100 clubs. uMinds runs Spanish speech-to-speech agents on campaigns of tens of thousands of calls. Sofratel and Bonnie&Car answer in French.
06 Questions
Multilingual voice agents, answered.
Which languages do your agents speak?
English, French, Dutch, Spanish and Portuguese, one language per agent.
Can an agent switch language during a call?
No. An agent speaks one language for the whole call. To serve several languages, run one agent per language.
Which speech-to-speech model sounds best in French, Spanish or Dutch?
Gemini 3.1 Flash Live has the best voice and feel of the field in every language we run, French, Spanish and Dutch included. In Spanish, Gemini 3.8 Live is close behind; in French and Dutch, the OpenAI and xAI voices trail behind.
Can it speak with a regional accent?
On speech-to-speech models, tone, accent and style are steered from the prompt. There is no regional setting: Spanish, for instance, is a single language.
Still have questions?
Contact us
More features
Real-Time Quality Monitoring
Every call recorded, transcribed, summarized and judged on your criteria, with an alert the moment it ends and a dashboard to review them.
Call Builder, API, Webhooks & Claude Skill
Draw your agent's call flow in the no-code Call Builder, then drive CallShift from your stack or from Claude: a REST API, tools your agents call mid-call, an end-of-call webhook and a Claude skill.
Next step
See it on your own calls.
Book a call with a voice AI engineer: we start from your real call flows and show you the agent, the architecture and the numbers that fit them.
1:1with a voice AI engineer, no slides
2architectures we ship in production: speech-to-speech and pipelines
1rollout plan to leave with: volume, pricing, telephony
~0% missed calls at Bonnie&Car, down from 55%. Read the case study