Revolutionizing IVR Systems: Attaching Voice Models to LLMs

by Dr. Phil Winder , CEO

Large language models (LLMs) are an established consumer facing technology, providing chat like functionality with artificial agents. However, many companies are now considering adding vocal capabilities, primarily to handle phone calls.

In this panel-based discussion, join Luke Marsden, Phil Winder, and other friends for a discussion about connecting voice models. We’ll discuss the idea, the complications, and the downstream effects.

Of particular interest is the unspoken expectation that when people call phone lines they expect to speak to humans. Even with the prevalence of script-based phone systems, do people still expect to speak to humans? Is it ethical to use an LLM it its place?

At the end of the discussion you will have a more grounded understanding of some of the key challenges involved with using voice models. This presentation will be high-level although we will discuss architectural concerns. It is suitable for all experience levels.

Download Slides

The following is an automated summary of the presentation.


Welcome to our latest blog post from Winder AI, where we delve into cutting-edge AI innovations. Recently, we hosted a webinar featuring Luke Marsden, an expert in cloud computing and AI, to discuss our new project, Helix, and the integration of voice models with large language models (LLMs). This blog post summarizes the key points and insights from the webinar.

Winder AI is a specialist AI agency dedicated to delivering high-quality AI projects. In our recent webinar, we were joined by Luke Marsden, a founder of several AI companies and a key player in cloud computing and AI ecosystems. Together, we explored the exciting potential of attaching voice models to LLMs and how this can revolutionize interactive voice response (IVR) systems.

Webinar Overview

The focus of the webinar was on integrating voice models with LLMs to enhance user experience in IVR systems. We discussed the current limitations of IVR systems and how LLMs can address these issues to provide a more natural and efficient interaction.

Current State of IVR Systems

IVR systems are widely used but come with several issues:

  • Common Issues:
    • Poor audio quality and outdated music.
    • Tedious navigation through multiple menu options.
    • Lack of flexibility in handling diverse queries.
  • Examples:
    • HSBC’s poor audio quality.
    • Use of manual button press codes to bypass menus.

Potential of LLMs in IVR Systems

LLMs offer significant advantages over traditional IVR systems:

  • Advantages:
    • Enable dynamic, fluid interactions.
    • Handle complex queries in natural language.
    • Improve user experience by bypassing rigid menu structures.
  • Challenges:
    • Need for careful prompting to reduce response latency.
    • Importance of concise system prompts for real-time interactions.

Privacy and Security Concerns

Privacy and security are paramount when integrating LLMs into IVR systems:

  • Data Privacy:
    • Compliance with GDPR and other regulations is crucial.
    • Concerns about sending data to third-party providers like OpenAI.
    • Importance of hosting models locally to ensure data security.
  • Regulatory Concerns:
    • European companies face restrictions on sending data to US-based cloud providers.
    • Need for private, secure deployments of LLMs.

Cost and Latency

Cost and latency are significant factors in the deployment of LLMs:

  • Cost:
    • Access to GPU resources can be expensive and complex.
    • Local deployment of GPUs can mitigate some of these costs.
  • Latency:
    • Local deployments can reduce network latency, improving response times.

Implementation and Ethics

Implementing LLMs in IVR systems involves several organizational and ethical considerations:

  • Organizational and Ethical Considerations:
    • LLMs should clearly identify themselves as AI to users.
    • Ensure availability of human support for complex or sensitive queries.
    • Avoid misuse of LLMs for scams or intrusive sales calls.

Demo Overview

During the webinar, we showcased a demo of integrating Vocode with Helix:

  • Vocode Integration with Helix:
    • Demonstrated low latency LLM interactions via voice.
    • Used DeepGram and Eleven Labs for speech-to-text and text-to-speech conversions.
    • Showcased LLMs’ ability to handle interruptions and maintain conversation flow.

Future with Multimodal Models

The future of AI lies in multimodal models, which promise richer and more accurate interactions:

  • Multimodal Models:
    • Offer a deeper understanding of user emotions and context.
    • Potential to reduce latency by integrating multiple modalities (voice, text, images) into a single model.

Recent Developments

  • Meta’s Chameleon Paper:
    • Introduces mixed modality models.
    • Indicates rapid progress in developing open-source multimodal models.

Final Thoughts and Q&A

In conclusion, LLMs have significant potential to revolutionize IVR systems:

  • LLMs can dramatically improve user interactions in IVR systems.
  • Local deployment of LLMs ensures better privacy, security, and cost management.
  • Continuous development and refinement of models are essential for broader adoption.

For more insights and future events, sign up to our newsletter. Stay tuned for more updates on how AI is transforming industries!

More articles

Build a Voice-Based Chatbot with OpenAI, Vocode, and ElevenLabs

Learn to create a chatbot using OpenAI, Vocode, and ElevenLabs for natural voice interactions. An example speech-to-text and text-to-speech system.

Read more

Practical Use Cases for Retrieval-Augmented Generation (RAG)

Join our webinar to explore Retrieval Augmented Generation (RAG) use cases and advanced LLM techniques to enhance AI applications in 2024.

Read more