IronHeart is not another voice AI. It is the operational layer that gives machines a persistent, emotionally-stateful presence — in robots, devices, companions, and digital humans.
Every other voice AI applies emotion as a filter — a pitch shift, a speed change, an EQ curve bolted on after synthesis. IronHeart bakes emotion into the generative process itself. The breath, the pause, the tremor — they emerge from within.
Emotion drives voice state: breathing rhythm, micro-pauses, tension, tremor, intimacy level. Not a filter applied after synthesis — these qualities are baked into the generative process from the ground up.
Each call or message is an event. Voice state and persona state — including intimacy level — persist across sessions. The AI remembers how it felt the last time. Identity is continuous, not reset.
Built-in RAG layer with external LLM-powered document rewriting for state-of-the-art retrieval. Configurable backend: OpenAI, Anthropic, Google, OpenRouter, or your own local model. KB from 1 GB to 15 GB.2,3
IronHeart runs offline. No cloud dependency at runtime. Deploy to edge hardware — on a robot, a kiosk, a companion device, or any embedded system. Designed for environments where latency and connectivity cannot be guaranteed.
Full API and SDK access. Integrate IronHeart into robots, devices, apps, digital humans, telephony systems, and AI companions. Voice becomes a controlled, programmable interface — not a black box.
IronHeart is infrastructure. It does not compete with your product — it powers the voice layer inside it, wherever that product lives.
Give robots a presence that matches their physical form. Stateful emotion lets a robot express fatigue, warmth, or urgency through voice — naturally.
Power the voice of avatars and synthetic personas that need consistent, emotionally coherent identity across long-running interactions.
Build companions that remember and evolve. Persistent persona state means the relationship grows — the AI recalls the emotional texture of past sessions.
Deploy stateful voice agents in call center infrastructure. Human-sounding, contextually aware, with built-in knowledge retrieval for live support.
Kiosks, smart speakers, in-vehicle systems. Run IronHeart offline with no latency penalty from cloud round-trips.
IronHeart is an API. Wire it into any system via REST or SDK. If your product needs voice — IronHeart is the layer beneath it.
Emotion and persona state are first-class citizens in the IronHeart API. Set them explicitly — or let the engine derive them from context and session history.
# Initialize with persona identity client = IronHeart(api_key="...") # Send a voice event response = client.speak( persona_id="aria-v2", text="I remember what you said last time.", # Emotion drives voice state — not post-processing emotion_state={ "valence": 0.3, # slightly subdued "tension": 0.6, # some held breath "intimacy": 0.8, # close, warm }, # Persist state to session persist=True, session_id="session_20260528" ) # Returns audio + updated persona state audio = response.audio_url state = response.persona_state
Credits are consumed per minute of voice call. Unused credits do not roll over.
1 Standard voice cloning requires 2–10 min of recording. Premium voice cloning: 1–4 hrs recommended for best quality.
2 Knowledge base LLM rewriting is optional but strongly recommended for state-of-the-art document retrieval accuracy.
3 LLM backend is configurable: OpenAI, Anthropic, Google, OpenRouter, or local models. Local model integration may require additional implementation support.
Tell us about your use case and we'll set up your API access. Robotics, digital humans, companions, enterprise telephony — we work with teams across the stack.