NVIDIA released Nemotron speech skills on October 13, 2026, saying they enable domain adaptation for speech recognition models. It is the company's first research update since the release of Nemotron 3.5 ASR Streaming 0.6B.
NVIDIA reported a 70.8% relative reduction in word error rate (WER) on the ATCO2 benchmark, measured on a two-hour human-annotated dataset. That compares with a 77.00% WER before adaptation.
Nemotron speech skills are built on NeMo speech and target domain-specific ASR adaptation, such as air traffic control. Availability begins with the release of the skills, initially for researchers and developers.
"A speech recognition model can handle everyday conversation and still struggle with a pilot reading back a clearance," said Francesco fciannella. The study highlights the need for staged adaptation to improve accuracy in specialized scenarios.
The announcement follows NVIDIA's release of the Nemotron 3.5 ASR model. NVIDIA framed the significance as a step toward improving real-world speech recognition performance in challenging environments.
NVIDIA did not say how the skills will perform on other domains, and the open question remains about generalizability across different use cases. The study plans to explore synthetic-text generation and acoustic profiling in future work.
Source: huggingface
