NVIDIA Nemotron3.5 Transcribe Author Benchmark - Context ASR - ContextASR-Bench - Fine - WER (%): leaderboard

NVIDIA Nemotron3.5 Transcribe Author Benchmark - Context ASR - ContextASR-Bench - Fine - WER (%) — NVIDIA Nemotron3.5Transcribe30BA4B own pinned BF16 model card; exact checkpoint c17667641642c4265ff2b7037dd101b7d4567634. Author evaluation, not automatically equivalent to native leaderboard protocols. ASR/translation/language ID reasoningoff temperature0; longform55minchunks5secoverlap. Audio/text reasoningon temperature0.6 top_p0.95 top_kdisabled seed0,20480completion budget including16384reasoning+1024finishing,65536context. No FP8/NVFP4 quality inheritance. Table 19 row 3. Native task/metric/cohort retained, no reconstruction of composites. Metric: WER (%) ↓. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B

Metric: WER (%) ↓. Source: huggingface.co. 1 model tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1Nemotron 3.5 Transcribe1.71

Interactive version: theaggregate.ai/benchmark?slug=nvidia-nemotron3-5-transcribe-author-benchmark-context-asr-contextasr-bench-fine-wer · How It Works · Data refreshed daily, snapshot 2026-11-16.