单个 600M 模型取代 40 个语言检查点
nvidia 的 nemotron 3.5 asr 在单个 600m 参数检查点中支持 40 种语言，具有原生标点和通过 cache-aware fastconformer 实现的低延迟流式传输。
nvidia introduced nemotron 3.5 asr, a 600m-parameter speech model designed to handle 40 languages and dialects within a single unified checkpoint, removing the need for per-language deployments or separate punctuation models. the architecture utilizes a cache-aware fastconformer encoder to process audio frames exactly once, significantly reducing latency and compute costs while maintaining accuracy compared to traditional buffered streaming approaches. it features native output of punctuated and capitalized text, and offers optional language id conditioning for either specified or automatic language detection in mixed-language scenarios.