October 2, 2026

We presented Contextual Earnings-22 at the datasets track.
In this paper, we introduce a first-of-kind dataset with golden-quality annotations to consolidate the benchmarking of contextual ASR (name and jargon recognition) across the industry. As a start, we benchmarked 10 systems across proprietary ASR APIs to open-source projects. Before our presentation at the conference, several teams across Nvidia, Hitachi and more had adopted our benchmark to report the resuils of their newly proposed systems.

Fun fact: The reason we describe the dataset annotations as golden-quality is that every single sample was manually reviewed, fixed and verified to set a 100% human baseline on the benchmark. In one extreme case, fixing an annotation required several minutes of cross-referencing person and company names on LinkedIn to disambiguate a person name where the original annotator marked "<inaudible>". Hundreds of similar cases were fixed during the review process.
In another track, we demoed the Argmax Pro systems paper, our frontier system for real-time speech-to-text with speakers and custom vocabulary on mobile devices.

Another fun fact: We scaled and tuned a 2-year-old paper from Nvidia as a core part of Argmax Pro. It was great to demo this scaled-up system to the Nvidia team (picture below) that authored the original work. They thought this particular algorithm was limited, moved on and proposed more advanced algorithms to solve custom vocabulary. Our work helped them realize the hidden potential of their earlier work.
Researchers visited the Argmax station at the expo to experience live demos of Argmax Playground (SDK) and pre-release consumer apps built on our SDK technology.

After a long but rewarding week, we explored the outskirts of Sydney:

The next major conference we are crashing is ICASSP 2027 with the following submitted work led by our summer current and past interns:
Consider reaching out to talk about you joinin Argmax to lead our next research project that leads to both a top-tier conference publication as well as immediate product impact.