NVIDIA Tso Tawm Nemotron 3.5 Transcribe Open-Weight Speech Model: Transcribes 40 Hom Lus, Nrog 463,062-Teev Audio Dataset

NVIDIA tau tso tawm Nemotron 3.5 Transcribe rau lub Kaum Hli 28, ib qho qauv hais lus qhib uas muab kev sau ntau hom lus, kev sau tus neeg hais lus, kev txhais lus, thiab kev teb cov lus nug hauv ib qho chaw kuaj xyuas. Nrog rau nws, lub tuam txhab tau tso tawm Granary v2, ib qho kev qhib cov ntaub ntawv hais lus thiab cov kav dej curation uas nws cov ntawv tshaj tawm pej xeem npog 463,062 teev ntawm lub suab thoob plaws 48 hom lus.
Raws li NVIDIA tshaj tawm, tus qauv no sau tau 40 hom lus, txhais 31 hom lus rau lossis los ntawm lus Askiv, thiab muab cov neeg hais lus sau ua 13 hom lus; nws tswj cov ntaubntawv kaw suab ntev txog ib teev thiab sau npe rau yim tus neeg hais lus. Tus qauv no siv 34-billion-parameter hybrid mixte-of-experts decoder (30B-A4B) uas ua kom kwv yees li 4.3 billion parameters ib lub token, thiab xa khoom hauv BF16, FP8, thiab NVFP4 quantized weights.
NVIDIA hais tias tus qauv no yog qib thawj zaug ntawm cov qauv qhib-qhov hnyav ntawm Hugging Face Open ASR English Leaderboard nrog 5.33% qhov nruab nrab ntawm cov lus yuam kev, thiab sib npaug rau thawj zaug ntawm cov qauv qhib-qhov hnyav ntawm Artificial Analysis Speech-to-Text ntawm 2.8%.
Cov qhov hnyav tau tso tawm raws li daim ntawv tso cai OpenMDW-1.1 rau kev siv ua lag luam lossis tsis yog kev lag luam, thiab muaj nyob ntawm Hugging Face, GitHub los ntawm NVIDIA NeMo Speech, NGC, thiab build.nvidia.com. Cov chaw xaus uas tau muab los ntawm Baseten, DeepInfra, thiab Simplismart, nrog rau kev sib tham sib qhia ntawm Hugging Face Spaces thiab ModelScope Studios.
Qhov kev tso tawm no yog tus qauv hais lus tshwj xeeb, sib txawv ntawm NVIDIA cov qauv Nemotron multimodal agent ua ntej. Kev sau cov lus hais yog ib qho thaiv rau kev siv super intelligence (SI), suav nrog cov ntawv sau sib ntsib, kev hu xov tooj rau cov neeg siv khoom, thiab kev sau cov lus hauv xov xwm.
Qhov chaw: GSM · Tebchaws Meskas (Qhov kev tshaj tawm lus Askiv · Qhov kev tshaj tawm lus Suav qub)
Duab: 图:英伟达位于加利福尼亚州圣克拉拉的总部大楼。CC BY-SA 4.0,经 Wikimedia Commons。 / Photo: NVIDIA headquarters in Santa Clara, California. CC BY-SA 4.0, via Wikimedia Commons.

