# NVIDIA reports lower air-traffic-control speech error after staged adaptation

> NVIDIA said staged training reduced Nemotron ASR's word error rate from 77% to 22.46% on a two-hour ATCO2 test set.

- Canonical URL: https://newsverum.com/en/news/nvidia-speech-adaptation-lowers-atc-error-october-2026
- Publisher: Newsverum
- Section: Technology
- Country: United States
- Published: 2026-10-15T19:00:45+00:00
- Updated: 2026-10-15T19:00:45+00:00
- Source review period: 2026-10-13
- Other editions: [Türkçe](https://newsverum.com/tr/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md), [العربية](https://newsverum.com/ar/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md), [Français](https://newsverum.com/fr/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md), [Español](https://newsverum.com/es/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md), [Deutsch](https://newsverum.com/de/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md), [Português](https://newsverum.com/pt/news/nvidia-speech-adaptation-lowers-atc-error-october-2026.md)

*The source does not state the event date; the period shown is the source review period.*

NVIDIA said its Nemotron 3.5 ASR Streaming model, adapted for air-traffic-control (ATC) speech, cut word error rate on a two-hour, human-annotated ATCO2 test from 77% to 22.46%. The company described that as a 70.8% relative reduction. On a general-English test, error rate rose from 3.50% before adaptation to 3.63% afterward.

The reported method first adapted broadly on noisy, machine-labelled ATCO2 recordings, then refined the model with limited human-labelled data; English examples were retained during training to limit losses on general speech. NVIDIA cautioned that the result does not guarantee the same gains in other domains. The post also introduced two coding-agent skills for planning, evaluating and reviewing ASR fine-tuning; installing those skills alone does not install the training software.

## Data

### Air-traffic-control speech recognition test

NVIDIA's Nemotron results on a two-hour, human-annotated ATCO2 test set.

Methodology: Word error rate counts insertions, deletions and substitutions against a reference transcript; lower is better. NVIDIA combined broad adaptation on noisy, machine-labelled ATCO2 data with a refinement stage using a small human-labelled Gold set. Values are comparisons reported in the company's 13 October 2026 post, not an independent replication.

| label | value | unit | period | basis |
|---|---|---|---|---|
| ATC word error rate before adaptation | 77 | percent | 2026-10-13 | observed |
| ATC word error rate after staged adaptation | 22.46 | percent | 2026-10-13 | observed |
| General-English word error rate before | 3.5 | percent | 2026-10-13 | observed |
| General-English word error rate after | 3.63 | percent | 2026-10-13 | observed |

Sources: <https://huggingface.co/blog/nvidia/fine-tune-asr-with-nvidia-nemotron-speech-skills> · Download data: [JSON](https://newsverum.com/api/v1/articles/nvidia-speech-adaptation-lowers-atc-error-october-2026/data/atc-asr-evaluation?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/nvidia-speech-adaptation-lowers-atc-error-october-2026/data/atc-asr-evaluation?lang=en&format=csv)

## Sources

- [NVIDIA, Speech Agent Skills study and tutorial](<https://huggingface.co/blog/nvidia/fine-tune-asr-with-nvidia-nemotron-speech-skills>) (Source checked: 2026-10-15; Published: 2026-10-13)

---

Cite the canonical URL above. [Editorial standards and corrections](https://newsverum.com/en/editorial)