How to Build an Empathic Support Bot with Hume AI EVI

Hume’s Empathic Voice Interface, usually called EVI, is a speech-to-speech loop that reads your tone and answers with an expressive voice. It is the part of Hume AI that product teams care about most. This guide builds a minimal support bot and flags where it breaks.

What is EVI, in one paragraph?

EVI is a real-time voice interface that combines speech recognition, emotion measurement, a language model and expressive text-to-speech in a single session. You speak, it understands both words and delivery, and it responds aloud with a tone it chose partly from your emotional signals.

How do you stand up a working session?

1. Create an API key and note your secret key for client-side token requests.
2. Open a WebSocket session with your config: system prompt, voice, and which emotion signals to expose.
3. Stream microphone audio in; stream the returned audio out for playback.
4. Read the transcript and emotion events as they arrive to drive any UI.

Keep secrets server-side. Issue short-lived tokens to the browser rather than embedding your key in client code.

How do you write the system prompt for a support role?

• State the role plainly: “You are a patient support agent for a billing product.”
• Cap reply length — long spoken answers feel lecturing.
• Define an escalation rule: when tension stays high for two turns, offer a human.
• Tell it never to state the user’s detected emotion out loud.

That last rule is the difference between attuned and unsettling.

What should you test before shipping?

Run scripted calls: an angry caller, a mumbler, someone who interrupts, and someone with a heavy accent. Watch latency, whether the bot talks over people, and whether emotion scores swing wildly on the same input. Also test a background-noise case, since phone audio is never studio-clean.

What are the hard costs and limits?

EVI bills by session duration, so a chatty bot that leaves the line open costs more than the quality win is worth. Free monthly credits cover demos only. Emotion inference is probabilistic, so treat signals as hints. And any production use storing emotional data must sort out consent first.

Quick Summary

• Hume EVI is a real-time speech-to-speech loop with built-in emotion awareness.
• Open a WebSocket session with a tight system prompt and keep keys server-side.
• Cap reply length and add an escalation rule for sustained tension.
• Test angry, mumbling, interrupting and accented callers plus noisy audio.
• EVI bills per session duration; free credits are demo-only and emotional data needs consent.