NVIDIA introduces Nemotron 3.5 ASR, a 600M-parameter streaming multilingual speech-to-text model supporting 40 language-locales with low latency, high accuracy, and built-in punctuation and capitalization. The article details how to fine-tune the model for specific languages, domains, or accents, showing significant WER reductions for Greek and Bulgarian as examples.

EVA-Bench Data 2.0 expands the enterprise voice agent benchmark from one domain to three: Airline CSM, Enterprise ITSM, and Healthcare HRSD, with 213 scenarios across 121 tools—a 4x increase. Scenarios are generated via the SyGra pipeline, manually reviewed, and validated by frontier models. Multilingual support is upcoming.

NVIDIA's team developed a task-seeded synthetic data generation pipeline that uses public task training splits as seeds to produce structured Q&A data for late-stage Nemotron model training. In a 100B-token continuation experiment, it improved MMLU-Pro by +1.8, average code by +1.9, commonsense understanding by +1.6, and GPQA by +11.1 while keeping math stable. The pipeline involves five stages: seed collection, record normalization, example generation, answer enrichment, and filtering. Transfer learning across task families enables the model to learn reusable behaviors.