跳到主要內容

Nemotron 3.5 Transcribe評測:Nvidia開源語音模型免費商用 | Nemotron 3.5 Transcribe Review: Nvidia's Open Speech Model

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 Nemotron 3.5 Transcribe評測:Nvidia開源語音模型免費商用

Nemotron 3.5 Transcribe 是 Nvidia 在 2026 年 10 月 28 日釋出的開源語音模型,一個 checkpoint 就能做轉錄、翻譯、辨識講者、回答語音內容的問題,而且授權允許商用、權重可自由下載。對比動輒要月付訂閱的雲端語音辨識 API,這次 Nvidia 的開源釋出是少數真的能自架、不綁供應商的選項。

Nemotron 3.5 Transcribe 是什麼?

這是一個 30B 參數的 Mixture-of-Experts 模型,但每次推論只啟動約 4B 參數,架構是混合式 Mamba–Transformer,語音編碼器搭配語言模型。重點是它「可被提示詞控制」:你可以用一般白話文要求「只要乾淨逐字稿」或「保留停頓與語助詞」,也能餵入醫療或產品專有名詞讓它優先辨識。

規格與效能

  • 支援 40 種語言轉錄、31 種語言可與英文互譯
  • 單一音檔可辨識最多 8 位講者,標出「誰在什麼時候說了什麼」
  • 單次處理音檔長度上限約 1 小時
  • 在 Hugging Face Open ASR 英文排行榜拿下開源模型第一名,平均字錯率(WER)5.33%
  • 在 Artificial Analysis 測試中 WER 低至 2.8%,與同類開源模型並列第一
  • 可在單張 Nvidia 顯卡上跑,從 Jetson 邊緣裝置到資料中心 GPU 都支援

跟雲端語音辨識服務比起來如何?

如果不想自架硬體,DeepInfra 等平台已經上架 Nemotron 3.5 Transcribe,報價約每分鐘 0.002 美元,大概是主流語音轉文字 API 行情的一半。對需要處理大量錄音(客服通話、會議紀錄、Podcast)的團隊,這個價差長期下來不小。

誰適合用、誰不適合

最大的限制是:這個模型目前只支援非即時(錄音檔)轉錄,不能拿來做直播字幕或即時語音助理這類低延遲場景。如果你要做的是會議錄音整理、Podcast 轉文字稿、多語言客服通話分析,這個模型很適合;但如果你要的是即時字幕或語音助理的即時應答,現階段還得搭配其他串流方案。另外權重雖然開源,實際部署仍需要一定的 GPU 資源與工程量,不是點兩下就能用的 SaaS 服務。

好不好用,試了才知道。


🇺🇸 Nemotron 3.5 Transcribe Review: Nvidia's Open Speech Model

Nemotron 3.5 Transcribe is the open-weight speech model Nvidia released on October 28, 2026. One checkpoint handles transcription, translation, speaker labeling, and Q&A about the audio content, and the license allows commercial use with freely downloadable weights. Compared to subscription-based cloud speech APIs, this is a rare open release you can actually self-host without being locked to a vendor.

What Is Nemotron 3.5 Transcribe?

It is a 30B-parameter Mixture-of-Experts model with only about 4B active parameters per inference pass, built on a hybrid Mamba-Transformer architecture pairing a speech encoder with a language model. The key feature is that it is promptable: you can ask in plain language for a clean transcript or a verbatim one with filler words kept, and feed it custom vocabulary, medical terms, product names, to prioritize during recognition.

Specs and Performance

  • Transcribes in 40 languages, translates between English and 31 languages
  • Labels up to 8 speakers per recording, who said what, and when
  • Handles recordings up to about 1 hour long
  • Ranks #1 among open-weight models on the Hugging Face Open ASR English leaderboard, with a 5.33% average word error rate
  • Hits 2.8% WER on Artificial Analysis, tied for #1 among open-weight speech-to-text models
  • Runs on a single Nvidia GPU, from Jetson edge devices to data-center chips

How Does It Compare to Cloud Speech APIs?

If you would rather not self-host, DeepInfra already offers Nemotron 3.5 Transcribe at roughly $0.002 per minute, about half what mainstream transcription APIs typically charge. For teams processing large volumes of audio, call center recordings, meeting notes, podcasts, that gap adds up fast.

Who Should (and Should Not) Use It

The biggest limitation: this model currently only supports non-streaming (recorded file) transcription, it is not built for live captioning or real-time voice assistants. If your use case is meeting transcription, podcast-to-text, or multilingual call analysis, it is a solid fit. But if you need real-time captions or low-latency voice interaction, you will still need a separate streaming solution. And while the weights are open, actually deploying it still takes real GPU resources and engineering effort, this is not a point-and-click SaaS tool.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code