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.
Source review period: 2026-10-13
Published: · Updated:
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.
- 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
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.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlyhuggingface.co ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- United States · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferredhuggingface.co ↗