Primary sourceIT之家
View sourceEvent ·Multi-source
NVIDIA 开源语音转录模型 Nemotron 3.5 Transcribe,支持 40 种语言
NVIDIA 于 10 月 28 日发布开放权重转录模型 Nemotron 3.5 Transcribe,总参数约 340 亿、激活约 43 亿,采用 MoE 架构。模型支持 40 种语言转录、31 种语言与英语互译,最长可处理 1 小时录音,并最多区分 8 名说话人。它在 Hugging Face Open ASR Leaderboard 英语组以 5.33% 平均词错误率居开放权重模型第一,在 Artificial Analysis 非流式词错误率指数以 2.8% 与 Voxtral Small Transcribe 并列开源首位。同步公开的 Granary v2 训练数据管线覆盖 463,062 小时音频、48 种语言。模型已在 Hugging Face 与 ModelScope 免费开放,采用 OpenMDW-1.1 许可证。
Current read
英伟达把转录、说话人区分和翻译打包成一个开源模型,40 种语言还能处理多人抢话,做会议纪要或字幕的可以试试。
11 sources67° AI heatLast updated 2026/10/29 07:31:29
Evidence
11 sourcesRelated source 1Artificial Analysis
View sourceRelated source 2kimmonismus
View sourceRelated source 3Hugging Face: Blog
View sourceRelated source 4marktechpost
View sourceRelated source 5rohanpaul_ai
View sourceRelated source 6vLLM
View sourceRelated source 7Decoder
View sourceRelated source 8TestingCatalog
View sourceRelated source 9techcrunch
View sourceRelated source 10pandaily
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.