COSCUP x UbuCon Asia 2026 was both my first overseas conference and my first UbuCon.
I joined the event as a volunteer at the Ubuntu Community Booth.
I had planned to introduce Ubuntu and its communities while demonstrating Kotonoha Interpreter, an Ubuntu-based side project I had been working on.
Things did not go as planned.
About Kotonoha Interpreter
Kotonoha Interpreter is an offline consecutive speech interpretation system designed for local deployment without reliance on cloud APIs.
It is not a simultaneous interpreter; rather, it waits for the speaker to conclude, performs speech recognition, translates the content into the selected language, and outputs the result as synthesized speech.
The system supports Korean, English, Traditional Chinese (Taiwan), and Japanese, offering 12 direct translation paths among these languages without using English as an intermediary.
The core processing pipeline is as follows:
Speech input → Speech recognition → Correction and translation → Speech synthesis
On the Jetson AGX Orin platform, Qwen3-ASR 0.6B serves as the primary speech recognition model. When recognition confidence is low, faster-whisper large-v3 is employed for verification. The automatic speech recognition (ASR) stage generates five candidate transcripts rather than a single output.
TranslateGemma 4B utilizes these candidates, recent conversational context, and a glossary to correct and translate the transcript in a single pass. The system streams output incrementally: each completed clause is forwarded to Qwen3-TTS 0.6B and rendered as soon as it is available.
For the demonstration setup, an NVIDIA Jetson AGX Orin was configured with the NVIDIA JetPack SDK, which includes Ubuntu-based Jetson Linux and NVIDIA’s AI development stack.
The complete source code is hosted in the Kotonoha Interpreter GitHub repository. Additional details regarding model configuration and the processing pipeline are provided in the system architecture documentation.
first photo shows the NVIDIA Jetson AGX Orin being set up and second photo shows the Kotonoha Interpreter WebUI.
What Went Wrong at the Venue
Before I left, the setup worked correctly; however, it failed upon arrival at the venue. An unforeseen issue with the on-device setup prevented initiation of the demonstration.
I dedicated significant time during the event to troubleshooting, including re-evaluating the configuration and runtime environment and attempting multiple recovery methods.
The issue could not be resolved within the allotted timeframe, leading to concern that the project might not be presented at all.
Codex, working hard alongside me to debug… with HBM Chips
While debugging, I kept the “We’re Under Maintenance” page displayed on the portable monitor.
Trying to Recover with an RTX A6000 Server
Unable to restore the on-device configuration promptly, I migrated the model backend to a server equipped with two NVIDIA RTX A6000 GPUs.
The Kotonoha Interpreter includes configurations for both Jetson devices and RTX A6000-class servers. On the A6000 platform, Voxtral Mini 4B Realtime serves as the primary automatic speech recognition (ASR) model, supplemented by faster-whisper large-v3 for ASR verification, TranslateGemma 12B for translation, and Qwen3-TTS 0.6B for speech synthesis.
Although the backend was successfully deployed on the server, the demonstration remained nonfunctional. The client was originally architected for local model services; modifying the backend address alone was insufficient to adapt the audio transport and processing pipeline for remote operation.
Consequently, the end-to-end interpretation pipeline could not be completed from the client through the remote server. By the conclusion of the event, neither the on-device nor the remote configuration was operational for a formal demonstration.
While the system functioned in a controlled environment at home, it failed at the venue, and the hastily assembled backup configuration was also ineffective. This outcome represented the most disappointing aspect of the event for me.
Continuing My Community Booth Work
While addressing technical issues during the project, I maintained my responsibilities at the community booth. I provided visitors with information about Ubuntu and UbuCon.
Additionally, I assisted my Ubuntu Korea LoCo colleague and associate, @choo121600 , in operating Ubu4Cut and guiding participants through the experience. Ubu4Cut is a photo booth initiative developed within the Ubuntu ecosystem, enabling attendees to take home a commemorative photograph from UbuCon.
I also promoted UbuCon Korea 2026 and highlighted the Ubuntu Korea LoCo community and its initiatives. Although my technical demonstration did not proceed as planned, I continued fulfilling my duties as a booth volunteer.
What I Learned
During the event preparation period, I concurrently served as an intern with the KREONET Team at the Korea Institute of Science and Technology Information (KISTI). Balancing this internship with the project resulted in insufficient development and validation time for Kotonoha Interpreter. Feature modifications and deployment adjustments continued shortly before the event, and comprehensive testing of model integration on the target hardware had not been completed.
Although a remote server was configured as a contingency, end-to-end validation of the client-to-backend pathway in a live environment was not performed. Consequently, the issues encountered at the venue were systemic rather than isolated. I should have initiated preparations earlier and completed full integration testing prior to the event.
In future iterations, I will begin preparations well in advance. For on-device deployment, I will conduct rigorous testing of model loading, memory utilization, audio input and output, and end-to-end latency on the target hardware.
For the remote configuration, validation must extend beyond backend startup to include the complete workflow—from audio transmission to translated speech playback. I will also maintain a clear distinction between development and stable demo builds. Before the next event, I intend to finalize a frozen demo release and ensure a prevalidated fallback configuration.
My objective is to resolve the identified issues, complete the outstanding validation, and deliver a more stable version of Kotonoha Interpreter.
My First Overseas Conference and First UbuCon
My inaugural overseas conference and first UbuCon did not proceed entirely as planned.
I regret being unable to demonstrate the Kotonoha Interpreter; the extensive time spent debugging on-site without resolution remains a memorable challenge. Nevertheless, the unsuccessful demonstration was not my sole takeaway.
Collaborating with colleagues at the community booth allowed me to present Ubuntu and the Ubuntu Korea LoCo community while engaging with representatives from other Ubuntu groups. Despite the technical issues, booth operations continued through collective support. Observing participants showcase their projects while assisting one another was instructive.
This experience underscored that rigorous preparation and real-world testing are as critical as project development itself.
A Day Trip
Outside the main event, I joined a day trip to experience Taiwan’s traditional tea culture. I learned how tea is made, hiked through the tea fields, picked tea leaves myself, and experienced the tea-making process firsthand.
As a side note, the last time I went hiking was in my first year of middle school. Hiking again after such a long time was exhausting, but it was also a lot of fun, haha.
Wrapping Up
I would like to extend my sincere appreciation to the event organizers, my colleagues at the community booth, the Ubuntu community, and all attendees who engaged with us.
Although I was unable to present a finalized version of Kotonoha Interpreter at this time, I am addressing the identified issues and will continue refining the project.
My objective for the next UbuCon is to present a more stable, comprehensive, and demonstration-ready iteration of Kotonoha Interpreter.
While the outcome diverged from my initial expectations, I established valuable new connections at this UbuCon, rendering the experience highly productive.
I truly hope we can all meet again at the next UbuCon Asia.
See you there!