Teleperformance — Voice First

Voice First

Client Request

Assistance chatbots have improved significantly, yet customer service still requires human support, especially through voice channels. How can we better support human assistants? This question led to the launch of the Voice First research project, a collaboration between Isagog and Teleperformance.

Challenges

Prototype development presented challenges in understanding spontaneous speech, representing knowledge, and performing automated reasoning. We addressed these by leveraging open-source reasoners and open language models.

A Cognitive Copilot to Inform Human Assistants

• "Listens" to the conversation by digitizing the customer’s voice (speech-to-text).
• "Understands" key concepts in real time.
• Retrieves information from a knowledge base: regulations, procedures, historical data.
• Reasons over a conceptual business model to provide suggestions to the human assistant: next best action, troubleshooting, applicable offers.

System Architecture

Development

Isagog’s reasoner for the Voice First project combines neural networks with symbolic reasoning and is built on three main components: Neural Parser – processes fragments of conversation, identifying and classifying key elements; Abductive Reasoner – interprets these elements by relating them to the customer’s context; Knowledge Base – holds business models and customer-specific information used for reasoning and decision-making.