英伟达发布开源语音转录模型 Nemotron 3.5 Transcribe

英伟达（Nvidia）发布了名为 Nemotron 3.5 Transcribe 的免费开源权重语音转录模型，旨在为开发者提供定制化的会议和医疗转录工具 1。该模型拥有 300 亿参数，支持多达 40 种语言的音频转录，并具备 31 种语言与英语互译的能力 1。在性能方面，英伟达宣称 Nemotron 3.5 Transcribe 在 Hugging Face Open ASR Leaderboard 等基准测试中优于竞争对手，能够处理长达 1 小时的录音并区分最多 8 个说话人 1。

为了配合新模型的训练与应用，英伟达同时发布了第二版开源语音数据集 Granary 1。技术规格上，该模型可在单个芯片上运行，兼容 Jetson AGX Thor、DGX Spark 以及 Blackwell 数据中心芯片等硬件平台 1。需要注意的是，Nemotron 3.5 Transcribe 仅适用于录制后的音频处理，不支持实时流媒体功能 1。

此次发布正值多家科技公司密集推出 AI 语音产品之际，竞争对手 Mistral 本月早些时候发布了 Voxtral 3，ElevenLabs 推出了 Realtime，而 Salesforce 则发布了 Launa-Omni 1。

Nvidia has released Nemotron 3.5 Transcribe, a new open-weight speech transcription model designed to handle audio processing for up to 40 languages 1. The model supports bidirectional translation between English and 31 other languages and can process recordings lasting up to one hour while distinguishing as many as eight speakers 1. This release arrives as the voice AI sector sees intensified activity, with competitors such as Mistral launching Voxtral 3 earlier this month, ElevenLabs introducing Realtime, and Salesforce releasing Launa-Omni 1.

Built on an architecture of 30 billion parameters, Nemotron 3.5 Transcribe is engineered to run efficiently on single-chip hardware solutions, including Jetson AGX Thor, DGX Spark, and Blackwell data center chips 1. Nvidia claims the model outperforms rivals in benchmark tests like the Hugging Face Open ASR Leaderboard, positioning it as a robust tool for developers creating custom transcription applications for business meetings and medical records 1. To support this initiative, Nvidia also unveiled the second version of its open-source speech dataset, Granary, though users should note that the model is restricted to pre-recorded audio and does not currently support real-time streaming 1.