Google has just made a notable breakthrough in interactive voice AI by officially introducing a new-generation conversational model duo: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. This is not simply a conventional speed-focused update, but a major leap in parallel reasoning capabilities combined with real-time visual processing.

According to official announcement from the Google Blog, the two new versions are specifically designed to address a long-standing bottleneck of previous virtual assistants: hesitation when they have to perform complex logical thinking while maintaining a natural conversation.

The core difference between the two Gemini 3.8 versions

Instead of bundling all features into one cumbersome package, Google clearly separates the two application goals to optimize infrastructure costs and operational performance for businesses:

  • Gemini 3.8 Live (Standard version): Optimized for large-scale deployment and cost efficiency. The model combines natural dialogue capabilities, direct visual data input (visual grounding), and automatic recognition and seamless switching between 97 languages in the middle of a conversation without interruption.
  • Gemini 3.8 Live Extended Thinking (Extended thinking version): Designed for tasks requiring deep multi-step reasoning. The unique feature is that AI can converse while silently processing complex logical algorithms, using natural filler phrases such as “Let me quickly check this part…” and narrating the processing progress for users to hear instead of remaining confusingly silent.

Gemini 3.8 Live

Key performance metrics and real-world benchmarks

Going beyond feature marketing claims, the Extended Thinking version has demonstrated superior capabilities through a series of rigorous standardized tests conducted by third parties:

Benchmark Score / Rate Actual meaning
Speech-to-Speech Quality Index (Artificial Analysis) 82.6 points (#1) The market's No. 1 natural voice conversation quality
Big Bench Audio 97.7% Extremely high accuracy in analyzing audio and context
τ-Voice Benchmark 68.6% Leading capability in automated task completion (Agentic Task)
τ-Voice Banking (Sierra) 35.1% Capability to handle demanding banking and financial operations

In ServiceNow's EVA-Bench test, which specifically measures voice assistants, the model pushed the Pareto Frontier to a new level by excellently balancing business accuracy with interaction smoothness.

3 breakthrough capabilities transforming the Voice Agent experience

What helps Gemini 3.8 Live move beyond the limitations of older generations of voice chatbots? The answer lies in 3 independent but smoothly coordinated operating mechanisms:

1. Calling tools in the background without interrupting the conversation

Previously, when you asked a bot: “Please check order #DPS-882 for me and reschedule the appointment to 3 PM tomorrow”, the bot would typically have to pause the conversation for a few seconds to call an API. With Gemini 3.8 Live, the model executes function calls and webhooks in the background while continuing to converse, announcing each processing step to keep the user experience completely seamless.

2. Ultra-fast visual input and response (Visual Grounding)

The model can “see” a camera or screen directly to provide assistance. For example, users only need to point the camera at a hand-drawn interface sketch on paper, and the AI will understand the structure and convert it into corresponding React source code in real time.

3. Transparent audio watermarking with SynthID

All audio generated by the model is embedded with a SynthID digital watermark that cannot be heard by the human ear but can be easily detected by moderation algorithms. This ensures transparency while minimizing the risk of voice deepfakes and impersonation fraud.

Gemini 3.8 Live

Application opportunities for Vietnamese businesses

For small and medium-sized enterprises (SMEs) in Vietnam, the launch of the Live API opens the door to advanced automation. Instead of maintaining costly telephone call center teams that can easily become overloaded during peak hours, companies can build intelligent Voice Agents:

  • Automated consultation and appointment booking: Connect through VoIP call centers to listen to and understand regional accents, look up pricing, and book appointments for clinics, spas, or auto garages.
  • Multi-channel operations automation: Combine the Live API with automation platforms such as n8n to synchronize live call data directly into CRM systems, Google Sheets, or Zalo ZNS.

If your business is looking for ways to optimize operating costs and improve customer experience through technology, explore the AI and n8n automation solutions at DPS.MEDIA to receive practical system architecture consulting tailored to real-world needs.

DPS.MEDIA JOINT STOCK COMPANY

Should change IP if listed > 2 times Address: 56 Nguyen Dinh Chieu, Tan Dinh Ward, Ho Chi Minh City

📞 Hotline / Zalo: 0961 545 445

✉ Email: marketing@dps.media

🌐 Website: https://dps.media

🏢 Tax identification number: 0318700500 | VCB: 1051160909

Ý kiến bạn đọc (9)

DPS SECURITY

Đăng nhập để bình luận

Đăng nhập nhanh 1-chạm bằng Google để chia sẻ ý kiến của bạn về bài viết.

Nguyễn Khánh Tùng
Nguyễn Khánh Tùng 16/09/2026

We are considering deploying a voice agent for our customer care department. We are not sure whether to choose Gemini 3.8 Live or the Extended Thinking version if we need both fast responses and order and policy information lookup?

Pham the hien
Pham the hien Experts 16/09/2026

If the task mainly involves scripted Q&A and needs to handle many calls at once, the standard version seems more suitable in terms of operating cost. For cases requiring multi-step reasoning or handling complex requests, try the Extended Thinking version first.

Kanon
Kanon 16/09/2026

I think we should test both on the same set of real-world scenarios, measuring response time, correct-answer rate, and cost per session. Don't just look at the published specifications and ignore the business's own call data.

nhimthanglong2023
nhimthanglong2023 16/09/2026

The ability to convert between 97 languages sounds impressive, but for Vietnamese, regional accents, slang, and how it handles customers who speak quickly still need further evaluation. A good demo does not necessarily mean it will work well in real-world use.

chungphamkimlongcom
chungphamkimlongcom 16/09/2026

The feature that lets it talk while calling tools in the background sounds quite convenient. But if the AI is checking data and the user interrupts or changes the question, how does the system handle it? Could it respond with the wrong context?

Nguyen Ha
Nguyen Ha Experts 16/09/2026

This is a key point to test thoroughly during real-world deployment. The AI must know how to prioritize new speaking turns, manage the status of running tasks, and confirm again before performing any action that affects data or transactions.

Thu
Thu 16/09/2026

Spot on. Natural conversation alone isn't enough. If we use it for a call center, we also care about correctly capturing requests, transferring to a staff member when needed, and not performing actions outside its scope on its own.

Van Buu Nguyen
Van Buu Nguyen 16/09/2026

Visual grounding can be useful for sales and technical support, but I want to know its accuracy when the camera shakes, in low light, or when the image contains multiple objects at the same time.

Hung Cuong Le
Hung Cuong Le Experts 16/09/2026

SynthID is a noteworthy approach for identifying AI-generated audio. However, businesses still need processes to verify content sources and clearly notify users, rather than relying solely on a single technical signal.

Ready to scale your revenue with DPS.MEDIA?

Partnering with 5,400+ SMEs to optimize marketing costs, elevate brand authority, and automate growth with 2026 AI technology.