An AI voice agent is a real-time spoken interface that talks to your users, takes action for them and runs on infrastructure you control. We build voice agents on two paths: a fast cloud path (ElevenLabs / Azure AI Foundry) for prototypes that need to ship in days, and a fully self-hosted open-source path (n8n + Whisper + Ollama + Piper) when sovereignty, cost or carbon constraints rule out hyperscalers. The second path is the one we used for Mother Earth AI – the self-hosted voice agent for climate communication that won the K3-Preis 2023.
Mother Earth AI is the clearest worked example. The project gives planet Earth a literal voice for climate communication. Their constraint: the system could not promote carbon emissions through hyperscaled AI providers. Sovereignty, autonomy and carbon-independence were non-negotiable.
We collaborated with the Mother Earth team on a fully self-hosted voice agent that runs on Ollama as the LLM platform and Open WebUI as the interface, with all components on the team's own infrastructure. The voice agent now serves two surfaces: the public website at mother-earth.ai, and a physical "Mutter Erde Telefon" – a Raspberry Pi-based phone installation that travels to museums, exhibitions and climate events, where visitors can pick up the receiver and have a spoken conversation with Mother Earth without an app or screen. The project won the K3-Preis 2023 für Klimakommunikation.
Most projects start on the cloud path – ElevenLabs for natural voice quality plus Azure OpenAI / GPT-4o via Microsoft AI Foundry for the language model. The cloud path ships an MVP in 1-2 weeks. Mother Earth AI is the counter-example we reach for when the project has a hard sovereignty, cost or carbon constraint that rules out hyperscalers – self-hosted takes longer to set up but gives you full control over data, cost and energy footprint. Learn more in our voice assistant guide.
Real-time conversation
Two delivery paths
Tool calling and action
A voice agent that delivers
Delivering self-hosted architectures for >3 years
Time to first MVP
1-2 weeks on the cloud path; longer for self-hosted, with a clear migration path from one to the other
Cutting-edge speech models
Parakeet, Whisper and Moonshine for speech-to-text; Kokoro and Piper for text-to-speech, in your language
Sovereignty by default
Self-hosted Ollama + open-source TTS / STT means audio never leaves your infrastructure if you don't want it to
Carbon-honest
Renewable-energy hosting where possible; transparent about the energy cost of voice models
1. Use case and architecture
2. Build the working agent
3. Deploy and operate
We bring a decade of impact-tech experience and over 160 AI projects since 2019. Through our free AI for Impact course, more than 100,000 people have learned how to use AI for the common good. We do not run inspiration days. We run scoping sessions and build engagements that ship, the way we have delivered AI for the organisations below:
Voice agents fail when they don't allow real-time interruption. Sprachnachricht-style request/response systems are easier to build but frustrate users. If real-time turn-taking matters, it has to be designed in from the start – not bolted on.
Voice quality is not solved. Cloud TTS providers like ElevenLabs still beat open-source TTS like Piper or Coqui on naturalness. Open-source is closing the gap fast but if voice quality is the hill to die on, the cloud path makes more sense.
Multilingual is uneven. Speech recognition and synthesis in major languages is excellent. In smaller languages and dialects it's still patchy. Test with your actual user group before committing.
Voice eats energy. Voice models are heavier than text models. We track and disclose the cost rather than hide it. For projects where carbon honesty matters – like Mother Earth AI – this shapes the architecture choice.
Cloud or self-hosted – which should we pick?
How long does an AI voice agent take to ship?
What languages do you support?
Can the voice agent take action?
Where does the audio go?
Can you build a physical voice installation like the Mutter Erde Telefon?
Tell us about the use case and the constraints. We will reply with a proposed architecture and a date, usually within a working day.
Simon Stegemann
Co-Founder and CEO
