Inspiration
International conferences bring together people who speak different languages, but real-time accessibility and translation solutions can be complex and expensive to deploy.
For the Nerdearla Vibeathon 2026, we wanted to explore a simple question: can we build an open-source system that turns live conference audio into accessible multilingual information while supporting multiple sessions at the same time?
That idea became Nerdearla Live Access.
What it does
Nerdearla Live Access is an open-source prototype for real-time conference transcription, bilingual translation, and session documentation.
The application provides three main workflows:
Conference Mode
- Continuous live conference sessions
- English to Spanish and Spanish to English processing
- Original transcription and translated text
- Multiple segments without losing the session history
- Manual speaker identification
- Session export to TXT, SRT, and VTT
Bilingual Conversation
- Alternating Speaker A and Speaker B turns
- English to Spanish and Spanish to English
- Shared chronological conversation history
- Original and translated text preserved for every turn
Operations / Concurrency
- Two independent Gemini Live sessions can run simultaneously
- Stage A and Stage B have independent controls and lifecycle
- One session can be stopped while the other continues processing
- The architecture can be extended through isolated workers for horizontal scaling
How we built it
The application was built during the Nerdearla Vibeathon 2026 using Next.js, TypeScript, React, the Google GenAI SDK, Gemini Live, and the Web Audio API.
Browser microphone audio is captured and converted to PCM audio for real-time processing.
A server-side API provisions short-lived credentials, keeping the permanent Gemini API credential outside the browser. The client then establishes its live session and processes streaming audio.
We also implemented a session-memory model that keeps structured transcript segments during the active browser session, including speaker, source language, target language, timestamps, original text, and translated text.
A technical glossary normalization layer helps maintain consistent terminology for conference-specific names such as Nerdearla, Vibeathon, Gemini, Devpost, OBS, and vMix.
Scalability
Each live pipeline is isolated as an independent session.
The Operations / Concurrency demo demonstrates this directly by running two simultaneous sessions and allowing one pipeline to stop while the other continues processing.
This session-worker pattern provides a foundation for horizontal scaling, where additional workers can handle additional conference sessions independently.
Challenges we faced
One of the main challenges was managing the lifecycle of real-time audio sessions: starting, pausing, resuming, switching speakers, finalizing transcript segments, and releasing microphone resources without losing session history.
We also worked on secure browser access to Gemini Live using short-lived credentials instead of exposing permanent API credentials to the client.
Another challenge was proving session isolation and concurrency rather than only describing scalability theoretically.
Accomplishments
During the Vibeathon we built and validated:
- Real-time transcription and bilingual translation
- Continuous conference session memory
- Bidirectional English/Spanish conversation
- Independent concurrent live sessions
- Technical terminology normalization
- TXT, SRT, and VTT export
- Secure short-lived browser credentials
- A deployed production demo
- A public open-source repository
What we learned
Building real-time multilingual systems is not only about transcription quality. Session lifecycle, latency, audio processing, security, state management, and concurrency are equally important when moving from a simple demo toward a conference-ready architecture.
What's next
Future development could add persistent server-side session storage, automatic speaker diarization, authentication and event management, additional languages, production monitoring, and integrations with broadcast workflows such as OBS or vMix.
The goal is to continue evolving Nerdearla Live Access from a Vibeathon prototype into a reusable accessibility layer for multilingual events.
Log in or sign up for Devpost to join the conversation.