NVIDIA releases Nemotron 3.5 Transcribe with open weightsMachine translation
NVIDIA 发布 Nemotron 3.5 Transcribe 开放权重语音转写模型
Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.
The model handles 40 languages and recordings up to one hour, using text instructions to produce verbatim transcripts, speaker-attributed output or translations. NVIDIA reports a 4.71% mean word error rate across nine public English speech recognition datasets, versus 4.94% for the next open-weights model in its report. It supports offline processing only, and Mandarin far-field meetings remain a weak spot.
This source permits summary display only.
Read at the original source