Description#
During my time as research associate in the Business Informatics program of HdM Stuttgart, I’ve designed and implemented XPerto, a voice assistant for multi-person conversations, particularly for scenarios like meetings and interviews.
It was part of a studentic research project, focusing on the acceptance of AI assistants and collaboration in professional settings. Unfortunately, I was not able to continue supervising this project and the upcoming user studies as my contract at the university ended before its completion.
The main features of XPerto include:
- Real-time interaction: Latency low enough to feel natural without sacrificing output quality.
- Multi-User Turn Detection: Automatic detection whether a response is needed or not based on the conversation context. Thus enabling multiple participants to interact with each other and/or the assistant.
- Transcript: As byproduct we are also able to supply a transcript of the interaction which can be further processed or summarized.
- Contextual Awareness: XPerto maintains context of the conversation to provide relevant responses across sessions.
Presentation at CONNECTIVAL'25#
We’ve had the chance to talk about and demo XPerto at the CONNECTIVAL conference in October 2025. It was a great success and sparked some insightful conversations about potential usecases to consider for further development. As requested I’ve attached the slides below, however they are German only (sorry for that).
Download the presentation (PDF)
Technologies#
- Pydantic: Used for settings management.
- Pipecat: The main framework used for building the Voice Assistant.
- Deepgram: Speech-to-Text service used for transcribing audio input.
- GPT 4o: Language model used for generating responses.
- ElevenLabs: Text-to-Speech service used for generating spoken responses.
Links#
Banner image © 2025 Anton Schneider
