Conversational Assistants that Converse with Each Other

## Overview

Deborah D and her team present the Open Voice Interoperability Initiative (OVON), a project within the LF AI and Data Foundation (formerly the Open Voice Network, founded in 2020) designed to enable conversational assistants—chatbots and voice assistants—to communicate with each other using open standards, similar to how web pages communicate via HTML and HTTP.

## The Problem

Millions of chatbots and voice assistants exist across mobile phones, smart speakers, and websites, hosted by governments, businesses, nonprofits, and educational institutions. However, each operates as an independent silo with its own expertise. When an assistant reaches the limits of its knowledge, users must abandon it and search for another, creating friction, duplication of effort, and implementation complexity across organizations with multiple legacy systems.

## The Vision

OVON proposes making conversational AI open and freely accessible like the web. Users could move seamlessly between assistants, enterprises could integrate legacy systems across different platforms, and innovators would have greater opportunity for interoperable functionality.

## Technical Architecture

OVON has developed three main specifications:

1. **Dialogue Events**: Describe the user request messages passed between assistants (e.g., "Do I need a visa to visit Estonia?"). 2. **Conversation Envelope**: The overall message format that invites assistants to join conversations, passes dialogue events, and handles goodbyes—uses JSON format with conversation IDs that persist across all participating assistants. 3. **Manifest**: Describes how assistants declare their capabilities so other assistants can determine if they can help (version 0.n nearly complete at time of talk).

The specifications are intentionally simple—only a few pages each—and support both speech and text interfaces. Internally, assistants can use any technology (rule-based, LLM-based, etc.); the standards only govern inter-assistant communication.

## Implementation and Demonstrations

OVON provides the **Sandbox**, an open-source implementation and testing platform. Version 2.0 adds automatic sequence diagram generation and transfer events. Demonstrations include:

- **Smart Library**: An LLM-based book information service accessible via phone (IVR) - **Bureaucrat**: An Estonian government services assistant - **Wizard**: A generic chat interface - **Order Assistant**: Demonstrated in IVR context providing order information

The Excal smart IVR example shows a practical phone-based system where callers can select different assistants and seamlessly transfer between them.

## Collaborations and Contributors

The primary collaboration is with the **government of Estonia**, which aims to create a unified portal for citizens to access all government services. Estonia serves as a real-world test case. Other contributors include implementers of the sandbox (Leah Barnes, Emma coin), specification authors (David Atwater), and various organizations developing independent implementations. The team emphasizes that contributors do the work while OVON provides guidance.

## Key Advantages

- **Maintainability**: Small, focused assistants are easier to maintain than monolithic systems; avoids catastrophic forgetting in ML retraining and concentrates developer expertise. - **Enterprise Integration**: Organizations can wrap existing assistants with OVON-compliant middleware without changing underlying systems (the Estonian approach). - **LLM Complementarity**: Since 96% of web content is not publicly available (proprietary, behind paywalls, personal data), large language models will never be comprehensive. Enterprise assistants must communicate internally; OVON standards provide a method. - **Openness**: Following web principles creates opportunities for entrepreneurs and allows users to benefit from distributed expertise.

## Current and Future Work

Dialogue events and conversation envelopes are being experimented with and polished. The manifest specification is nearing completion. Discovery mechanisms—standards for finding appropriate assistants (analogous to web search engines)—are just beginning, requiring specifications for communication with discovery agents and for passing conversation context. Multimodal data support is identified as a future need.

## Resources

All work is available in a public GitHub repository, with webinars on YouTube covering specifications, the sandbox, and the Estonia collaboration.