STELLA v1.1.0: making voice feel like conversation rather than a queue

This started as a LinkedIn post announcing the release. Here is the longer version.

STELLA v1.1.0 is out. STELLA is our open-source infrastructure for running studies with your own LLM-based conversational agents. If you have not met it yet, our paper introduces the platform: STELLA: An Open-Source Platform for Auditable Voice-First Conversational Health Coaching Agents, published at the ACM Conference on Conversational User Interfaces.

This release is mostly about one thing: making voice feel even more like conversation rather than a queue.

Faster voice

• First audio: from about 1.6 seconds down to 900 milliseconds, and it no longer scales with sentence length.
• Time-to-first-audio from TTS: 30% lower.
• Gap between sentences: from 240 milliseconds down to 160 milliseconds.

One agent instead of two

Because of these latency gains, we are consolidating onto a single agent and retiring the light one. You now get the control and oversight of the main agent at the speed of the lightweight one.

Build your own agent

If the default agent does not fit your needs, STELLA's Python SDK lets you build your own and deploy it to the same infrastructure. Implement process(), yield outputs, and STELLA handles the rest. STELLA has no opinions about how your agent thinks. Your logic stays yours, and STELLA just contributes what has already been tested on our end.

Try it

Head over to GitHub and check it out: github.com/c4dhi/STELLA.