Nvidia 近期推出了开源语音识别模型 Nemotron 3.5 Transcribe，该模型具备开放权重且支持提示功能，能够处理40种语言的转录任务 1。在语言处理能力方面，其中31种语言支持双向翻译至英语，另有13种语言实现了说话人分离功能 1。这一发布旨在为开发者提供高精度的多语言语音处理工具，并配套推出了 Granary v2 数据处理管道及微调套件以辅助特定领域的适配 1。
在性能表现上，Nemotron 3.5 Transcribe 在 Hugging Face Open ASR Leaderboard 的英语赛道中凭借 5.33% 的平均词错误率（WER）位居开源模型榜首 1。此外，在 David AI 的多语言说话人分离基准测试中，该模型的最低错误率为 18.26% 1。为了支持模型的训练与优化，Granary v2 数据集涵盖了来自16个源语料库、涉及48种语言的总计 463,062 小时音频数据 1。
除了通用场景的高性能，该模型在垂直领域也展现出显著优势。案例研究显示，经过微调后，其在医疗领域的保留说话人词错误率相比最佳提示基线降低了 45.8%，从 22.44% 下降至 12.17% 1。目前，用户可以在 RTX Spark、DGX Spark 等本地硬件或 Blackwell 数据中心 GPU 上运行该模型 1。
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight, promptable automatic speech recognition (ASR) model that supports transcription for 40 languages 1. The system enables translation between English and 31 of these languages, while also offering speaker diarization capabilities for 13 specific languages 1.
The model ranks first among open-source models in the English track of the Hugging Face Open ASR Leaderboard with an average word error rate (WER) of 5.33% 1. It also achieved the lowest error rate of 18.26% in David AI’s multilingual speaker diarization benchmark 1.
To support domain-specific adaptation, NVIDIA introduced Granary v2, a data processing pipeline built on a dataset containing 463,062 hours of audio across 48 languages from 16 source corpora 1. In a medical field case study, fine-tuning using this pipeline reduced the WER by 45.8% compared to the best baseline prompts while retaining speaker information, dropping the rate from 22.44% to 12.17% 1. The model is compatible with local hardware such as RTX Spark and DGX Spark, as well as Blackwell data center GPUs 1.
评论
还没有评论，欢迎留下第一条。