跳到主要內容

MAI-Transcribe-2評測:語音轉文字最準,但比對手貴 | MAI-Transcribe-2 Review: Accurate, But Pricier Voice AI

By Kit 小克 | AI Tool Observer | 2026-10-03

🇹🇼 MAI-Transcribe-2評測:語音轉文字最準,但比對手貴

Microsoft 在 2026 年 10 月 1 日同時推出三款語音 AI 模型,其中最受關注的是即時語音轉文字模型 MAI-Transcribe-2-Streaming,加上文字轉語音的 MAI-Voice-2.1 與 MAI-Voice-2.1-Flash。重點很簡單:MAI-Transcribe-2 在跑分拿下目前最準的即時語音轉文字成績,但換算下來的價格比多數對手貴,這篇用實際數字拆給你看值不值得換。

MAI-Transcribe-2-Streaming 準確度真的最高嗎?

是,根據 Artificial Analysis 的跑分,MAI-Transcribe-2-Streaming 在 38 個模型裡拿下準確度第一:最終字錯率只有 2.5%,初步結果(partial)錯率 2.8%,收到音訊後約 100 毫秒就能吐出第一版文字,語音結束後 0.13 秒內定稿,支援 60 種語言並能自動偵測。

MAI-Transcribe-2 多少錢?比對手貴多少?

MAI-Transcribe-2-Streaming 到 2026 年底前的優惠價是每小時音訊 0.54 美元,換算每千分鐘 9 美元。這個價格跟 Google Gemini 3.5 Transcribe Live 打平,但明顯比下面這些對手貴:

  • ElevenLabs Scribe v2 Realtime:每千分鐘 6.5 美元
  • Deepgram Flux:每千分鐘 6.5 美元
  • Cartesia Ink-2:每千分鐘不到 4.5 美元,比 Microsoft 便宜超過一半

換句話說,MAI-Transcribe-2 準確度最高是事實,但如果你的應用對字錯率沒那麼敏感(例如會議記錄、客服逐字稿),Deepgram 或 Cartesia 的性價比明顯更好。

MAI-Voice-2.1 跟 Flash 版差在哪?

MAI-Voice-2.1 走高擬真路線,每百萬字元 22 美元,盲測中有 50.3% 聽眾覺得它跟真人錄音一樣自然甚至更自然;MAI-Voice-2.1-Flash 主打低延遲、高流量場景,150 毫秒端到端延遲、推論速度快 55%,價格只要每百萬字元 15 美元,比同類模型便宜約 60%。兩者都支援用幾秒參考音訊複製聲音,且內建「同意」防護機制。

開發者現在能用嗎?

三款模型都在 Microsoft Foundry 公開預覽中,可用 Azure Speech SDK 或相容 OpenAI Realtime 規格的 WebSocket API 呼叫,語音模型也已上架 Vercel、OpenRouter 和 Azure Voice Live,LiveKit 整合即將推出。

好不好用,試了才知道。


🇺🇸 MAI-Transcribe-2 Review: Accurate, But Pricier Voice AI

On October 1, 2026, Microsoft launched three new voice AI models at once: the real-time speech-to-text model MAI-Transcribe-2-Streaming, plus two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The headline: MAI-Transcribe-2 just took the top accuracy spot on a major benchmark, but it costs more than most rivals — here's the actual math on whether it's worth switching.

Is MAI-Transcribe-2-Streaming really the most accurate?

Yes. On the Artificial Analysis benchmark, MAI-Transcribe-2-Streaming ranks #1 for accuracy among 38 models: a 2.5% final word-error rate and 2.8% partial-rate, with first hypotheses arriving around 100ms after audio starts and a final transcript locked in 0.13 seconds after speech ends, across 60 auto-detected languages.

How much does MAI-Transcribe-2 cost compared to rivals?

MAI-Transcribe-2-Streaming runs an introductory $0.54 per hour of audio through year-end — about $9 per 1,000 minutes. That ties Google's Gemini 3.5 Transcribe Live, but it's noticeably pricier than:

  • ElevenLabs Scribe v2 Realtime: $6.50 per 1,000 minutes
  • Deepgram Flux: $6.50 per 1,000 minutes
  • Cartesia Ink-2: under $4.50 per 1,000 minutes — less than half Microsoft's price

Top accuracy is real, but if your use case — meeting notes, support transcripts — can tolerate a slightly higher error rate, Deepgram or Cartesia likely offer better value per dollar.

What's different between MAI-Voice-2.1 and the Flash version?

MAI-Voice-2.1 targets maximum realism at $22 per million characters — in blind tests, 50.3% of listeners rated it as natural or more natural than actual human recordings. MAI-Voice-2.1-Flash is built for high-volume, latency-sensitive use: 150ms end-to-end latency, 55% faster inference, and roughly 60% cheaper at $15 per million characters. Both support voice cloning from a few seconds of reference audio, with built-in consent safeguards.

Can developers use it today?

All three models are in public preview on Microsoft Foundry, callable via the Azure Speech SDK or an OpenAI Realtime-compatible WebSocket API. The voice models are also live on Vercel, OpenRouter, and Azure Voice Live, with LiveKit integration coming soon.

You won't know until you try it.

Sources / 資料來源

常見問題 FAQ

MAI-Transcribe-2-Streaming 是什麼?

是 Microsoft 在 2026 年 10 月發表的即時語音轉文字模型,主打低延遲與高準確度,收到語音後約 100 毫秒就能吐出初步文字結果。

MAI-Transcribe-2 比 Deepgram、ElevenLabs 準嗎?

根據 Artificial Analysis 跑分,MAI-Transcribe-2-Streaming 以 2.5% 字錯率排名第一,但價格每千分鐘 9 美元,比 Deepgram Flux、Cartesia Ink-2 等對手貴。

MAI-Voice-2.1 可以拿來複製別人的聲音嗎?

技術上支援用幾秒參考音訊複製聲音,但 Microsoft 內建了同意(consent)防護機制,限制未經授權的聲音複製。

現在可以用這些模型嗎?

可以,三款模型都已在 Microsoft Foundry 公開預覽,透過 Azure Speech SDK 或 WebSocket API 呼叫,也能在 Vercel、OpenRouter 上使用。

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