Google 推出 Gemini 3.8 Live 雙模型,語音 AI 登頂基準測試,支援「邊講邊思考」
作者 KOCPC 編輯團隊
2026 年 9 月 17 日
屬於 AI 趨勢及相關新聞, 最新科技新聞
9月15日,谷歌宣布推出两款新的语音模型:Gemini 3.8 Live和Gemini 3.8 Live Extended Thinking,旨在实现实时语音代理应用。其中,Extended Thinking版本在Artificial Analysis的Speech-to-Speech Quality Index上获得了82.6分,排名第一。这两款模型现已通过Gemini API和Google AI Studio提供给开发者,并强调了在对话中执行任务的能力。该模型可以在后台调用API或工具,同时继续向用户输出语音响应,而不会打断对话流程。
兩檔 Gemini 3.8 Live 模型的定位差異
Gemini 3.8 Live:定位為 speech-to-speech 模型,專注於大規模部署和成本效率,在 Speech Agent Arena 的用戶偏好度排名第二。核心功能包括:
異步函數調用:在後台執行 API 和工具調用,同時持續向用戶流式傳輸語音回應。
實時視覺上下文:將實時視覺輸入納入會話上下文,使語音代理可以理解用戶「說」和「看到」的內容。
字母數字精確度:準確解析確認碼、索賠號碼和技術規範等字母數字資訊。
97+ 種語言支援:對話中可自動檢測和切換語言,無需手動配置。
Gemini 3.8 Live Extended Thinking:專為高度複雜任務設計,具備更深層的推理能力。它支援「可配置的思考」,可以在後台處理多步推理,同時使用口語提示(例如「讓我核對一下」)讓對話無縫繼續,並即時發佈任務進度。
Gemini 3.8 Live 基準測試得分
Google 發布的基準數據顯示:
Artificial Analysis Speech-to-Speech Quality Index:Extended Thinking 以 82.6 分奪得首位,而標準版 3.8 Live 排名第二。
τ-Voice 代理任務完成率:Extended Thinking 達到 68.6%
Sierra τ-Voice-banking: 35.1%
Big Bench Audio 推理能力:97.7%
ServiceNow EVA-Bench:在準確性和會話品質的天平上推進了 Pareto 前線。
在一篇博客文章中,谷歌DeepMind的高级工程师Tom Ouyang表示,这两款模型相较于之前的Live模型是一个“巨大的飞跃”,并提供了一个更优化的替代方案,用于克服级联架构的局限。
Gemini 3.5 Transcribe:同步整合語音轉文字模型
Google 還將上個月發布的 Gemini 3.5 Transcribe 整合到了同一個開發者工具包中。這款 speech-to-text 模型支援 85+ 種語言,在流式模式下的字詞錯誤率 (WER) 為 4.0%,非流式模式下為 2.6%。
特殊功能包括:
自動語言切換:可以在句內和句間自動處理語言切換,無需手動配置。
自定義詞彙表偏好:可以傳入一個包含 1,000 個技術術語的列表,引導語音識別優先識別特定領域的詞彙。
智能轉錄模式:自動去除填充詞,糾正口誤,並輸出結構化、可直接閱讀的轉錄文字。
3.5 Transcribe 還可以通過 Interactions API 處理長達 1 小時的音頻文件,並支援結構化的時間戳和說話者標籤。
定價和合作夥伴
Gemini 3.8 Live 系列的定價為每分鐘音頻輸入 $0.005(約 NT$0.16),音頻輸出為 $0.018(約 NT$0.59),換算成 Token 價格大約是每百萬 Token 輸入 $3,輸出 $12。
在合作伙伴方面,Agora、Fishjam、LiveKit、LangChain、Pipecat、Vercel 和 Vision Agents 等平台现在都支持 Live API,处理实时媒体流基础设施。企业合作伙伴包括 Salesforce、ServiceNow、Genspark 和 Lumeris。
谷歌还提供了完整的语音工具包,包括 Gemini 3.5 Live Translate(70 多种语言的语音到语音翻译)、Gemini 3.1 Flash TTS(语音生成)和 Lyria 3.5(音乐生成),均可通过 Gemini API 访问。
数据来源
来源:KOCPC 中文
标签:Gemini、Gemini 3.5 转录、Gemini 3.8 直播、Google
Google launches Gemini 3.8 Live dual models, with its voice AI topping benchmarks and support for “thinking while speaking.”
by KOCPC Editor
September 17, 2026
in AI Trends and Related News, Latest Technology News
On September 15, Google announced two new voice models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, targeting real-time Voice Agent applications. The Extended Thinking version scored 82.6 on Artificial Analysis’ Speech-to-Speech Quality Index, ranking first. Both models are available to developers through Gemini API and Google AI Studio, highlighting the ability to execute tasks during a conversation. The models can call APIs or tools in the background while continuing to output voice responses to users, without interrupting the conversation flow.
Positioning differences between the two Gemini 3.8 Live models
Gemini 3.8 Live: Positioned as a speech-to-speech model prioritizing large-scale deployment and cost efficiency, it ranked second in user preference in the Speech Agent Arena. Core capabilities include:
Asynchronous function calls: Execute API and tool calls in the background while continuously streaming voice responses to the user.
Real-time visual context: Incorporate real-time visual input into the conversational context, enabling voice agents to understand what users “said” and “saw.”
Alphanumeric precision: Accurately parse alphanumeric information such as confirmation codes, claim numbers, and technical specifications.
Support for 97+ languages: Automatically detects and switches languages during conversations, with no manual setup required.
Gemini 3.8 Live Extended Thinking is designed for highly complex tasks and has deeper reasoning capabilities. It supports “configurable thinking,” can process multi-step reasoning in the background while using spoken cues (such as “Let me verify…”) to keep the conversation uninterrupted, and announces task progress in real time.
Gemini 3.8 Live benchmark scores
The benchmark data released by Google shows:
Artificial Analysis Speech-to-Speech Quality Index: Extended Thinking takes first place with 82.6 points, while the standard version 3.8 Live comes in second.
τ-Voice agent task completion rate: Extended Thinking reaches 68.6%
Sierra τ-Voice-banking: 35.1%
Big Bench Audio reasoning capability: 97.7%
ServiceNow EVA-Bench: Advanced the Pareto Frontier in balancing accuracy and conversational quality.
In a blog post, Google DeepMind Principal Engineer Tom Ouyang said the two models deliver a “step change” over the previous Live models and offer a more streamlined alternative to cascaded architectures.
Gemini 3.5 Transcribe: Synchronized integration of speech-to-text models
Google has also integrated Gemini 3.5 Transcribe, launched last month, into the same developer toolkit. This speech-to-text model supports 85+ languages, with a word error rate (WER) of 4.0% in streaming mode and 2.6% in non-streaming mode.
Special features include:
Automatic code-switching: Language switching within and between sentences is handled automatically, with no manual configuration required.
Custom vocabulary preference: Allows passing in a list of up to 1,000 technical terms to guide speech recognition to prioritize recognition of domain-specific vocabulary.
Smart Transcription Mode: Automatically remove filler words, correct verbal slips, and output structured, directly readable transcript text.
3.5 Transcribe can also process audio files up to 1 hour long via the Interactions API, with support for structured timestamps and speaker labels.
Pricing and Partners
Pricing for the Gemini 3.8 Live series is $0.005 per minute for audio input (about NT$0.16) and $0.018 per minute for audio output (about NT$0.59), which converts to approximately $3 per million Tokens for input and $12 per million Tokens for output.
On the partner side, platforms such as Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, and Vision Agents now support Live API, handling real-time media streaming infrastructure. Enterprise partners include Salesforce, ServiceNow, Genspark, and Lumeris.
Google also offers a complete voice toolkit, including Gemini 3.5 Live Translate (speech-to-speech translation for 70+ languages), Gemini 3.1 Flash TTS (speech generation), and Lyria 3.5 (music generation), all accessible through Gemini API.
Data source
Source: KOCPC Chinese
Tags: Gemini, Gemini 3.5 Transcribe, Gemini 3.8 Live, Google
