The Intent Engine is the core intelligence layer of Seminara. It is responsible for parsing user speech in real-time and deciding how Aura should react.
The Finite State Machine (FSM)
Seminara does not use a generic "chatbot" loop. Instead, it relies on a strict 6-phase Orchestrator FSM (Initialization, Narration, Interruption, Evaluation, Generation, Resumption).
This ensures that the presentation maintains structural integrity. The session will always progress toward the Call to Action, regardless of how many times the attendee interrupts.
Handling Interruptions
When an attendee speaks:
- Voice Activity Detection (VAD): Aura immediately stops speaking to listen.
- Transcription: Our ultra-low latency models transcribe the speech in milliseconds.
- Intent Classification: The Intent Engine determines if the user made a COMMAND (e.g., "Stop"), a QUERY (e.g., "How much does it cost?"), or noise.
- Contextual Generation: The LLM generates a grounded response using the Knowledge Base.
- Resumption: Aura answers the question, then seamlessly resumes the presentation from the exact bookmark where she was interrupted.
Frequently Asked Questions (FAQ)
Q: What happens if two people speak at once?
A: Currently, Seminara sessions are optimized for 1-on-1 or small group interactions where turn-taking is managed by the host agent. The VAD will trigger on the loudest distinct speech.
Q: Can the Intent Engine handle multiple languages?
A: Currently, Seminara officially supports English (US, UK, AUS). While the transcription model can detect other languages, Aura's generation and voice models are highly optimized for English.
Q: Does the Intent Engine support custom hotwords?
A: No, attendees don't need to say "Hey Aura." The engine uses continuous Voice Activity Detection (VAD) to naturally detect when an attendee is addressing the host.
Q: How fast is the Intent Engine's response time?
A: Our system is optimized for ultra-low latency. The round-trip time from the end of the user's speech to the start of Aura's verbal response is typically under 800 milliseconds.