14 OCTOBER 2026
Adapting speech recognition for Hindi–English conversations.
Dataset preparation, supervised experiments and adapted-checkpoint comparisons. Comparative findings under internal review.
RESEARCH NOTES & WORKING PAPERS
The methods. The prototypes. The next questions.
Explore the methods behind our speech, vision and agentic work: completed speech studies, ongoing visual-model investigations, video prototypes and workflow designs.
The papers cover speech adaptation, speaker attribution, dialogue, video repair and agentic workflows, with ongoing NVIDIA, Google and Apple model experiments on the cloud Mac. Comparative speech findings are under review.
14 OCTOBER 2026
Read the note ↗
14 OCTOBER 2026
Which speaker said which words?
Anonymous speaker attribution, segmentation adaptation and separate assessment of transcription errors.
14 OCTOBER 2026
Interruptions, corrections and spoken replies.
Moshi feasibility, controlled interactions and the requirements for further adaptation.
14 OCTOBER 2026
One character across a sequence.
Prototype failures, scene constraints, selective repair and proposed comparisons with regeneration.
14 OCTOBER 2026
Designing agentic workflows.
Research automation and a proposed architecture for sources, tools, review and recovery.
14 OCTOBER 2026
Testing models and workflows.
Speech evaluation methods and further plans for model, workflow and video comparisons.
DISCUSS THE RESEARCH
Questions
and collaboration.
Bring a speech or vision task, a failure case or a different evaluation perspective. We welcome technical discussion and focused research collaboration.
Discuss the research ↗