# 🧰 Builders building with Gemini Audio

## Executive summary

The video showcases various applications of Google's latest AI models, including Gemini Audio, Gemma, and Lyria. Discussions highlight building powerful local applications using Gemma's open-weight capabilities, creating music with Lyria, and developing real-time voice accessibility tools. Key technical features discussed include the Gemini 3.8 client connecting via WebSockets, the utility of Gemma models in environments lacking Wi-Fi, and advanced features like 'Extended Thinking' for long-running background processes.

## Key takeaways

- Gemini 3.8 and WebSockets: The Gemini 3.8 client can connect to Gemini 3.8 running in the cloud using WebSockets, enabling real-time interaction.
- Gemma for Offline Use: Gemma models are highlighted as ideal for environments where a stable Wi-Fi connection may not be available, facilitating local application development.
- Advanced AI Features: The 'Extended Thinking' feature allows for long-running processes to execute in the background, even when the primary conversation is unrelated to the asynchronous tool call.
- MAviS Benchmarking: Google released the MAviS Bench, a benchmark used to compare speech-to-speech models, where the featured models performed strongly.

## Technical details

- Gemini Audio & Gemini 3.8: The Gemini 3.8 client connects to the cloud version of Gemini 3.8 using WebSockets for real-time audio processing.
- Gemma Models: Gemma models are emphasized for their ability to function in environments without Wi-Fi, supporting local, powerful applications.
- Real-Time Voice Accessibility: Demonstrations include using Gemini Live or Gemma models to provide voice descriptions of visual information for individuals with visual impairments.
- Lyria Models: Lyria models are discussed in the context of advanced music creation, viewed as a collaborative 'partner' for the user.

## Practical implications

- Build engineers can leverage Gemma's open-weight nature for developing robust, offline-capable applications.
- Integration of Gemini 3.8 via WebSockets allows for building low-latency, real-time voice and audio tools.
- The 'Extended Thinking' feature provides a mechanism for managing complex, multi-stage background processes within an application's user experience.

## Topics

Generative AI, Large Language Models (LLMs), Audio Processing, Edge Computing, WebSockets, Gemini Audio, Gemma

Source: https://www.youtube.com/watch?v=LJ2v4-FvPcA