AI Today: Gemini 3.8 TTS and robotics infrastructure

State of knowledge as of 2026-09-24. Source material gathered on 2026-09-24.

Today brought a series of announcements in the fields of speech synthesis, conversational systems, and hardware infrastructure. Google introduced advanced Gemini 3.8 TTS models, while OpenAI focused on mental health safety and education in Asia. Simultaneously, companies like Ringg are deploying efficient GPT-5.6 models for customer support, and Microsoft is optimizing inference infrastructure for physical robotics by offloading computing to the cloud.

Key takeaways today (TL;DR)

• Gemini 3.8 TTS: Google unveiled new text-to-speech models supporting over 100 languages with advanced audio control features.
• MentalHealthBench: OpenAI created an expert-driven benchmark to evaluate the safety of AI responses in mental health conversations.
• Ringg: The company deployed GPT-5.6-based AI agents, reducing customer service costs by 90%.
• Grab: OpenAI and Grab launched the GO Forward with AI educational program for 30,000 partners in Southeast Asia.
• Microsoft: Researchers developed a method to increase robot efficiency by offloading AI computations from onboard units to external infrastructure.

Google introduces Gemini 3.8 TTS models

According to an announcement on September 23, Google introduced two new text-to-speech models: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. These systems offer advanced voice synthesis capabilities, supporting over 100 languages and dialects.

The Gemini 3.8 Flash TTS model is designed for character creation and advanced sound design. It allows for the creation of new voices from scratch using natural prompts and provides detailed control over performance parameters such as pace, emotion, and speaking style. Meanwhile, Gemini 3.8 Flash-Lite TTS is optimized for performance and cost, making it a suitable solution for large-scale content creation such as dubbing or voice agents.

Users can choose from a library of over 2,000 pre-made voices or replicate an existing voice timbre using a 30-second sample. The models also enable precise dialogue control, including multi-speaker support in a single scene and the addition of natural insertions like laughter or breaths.

For security, Google introduced voice owner consent verification mechanisms and SynthID watermarking to digitally tag generated audio. The models secured top positions in Hume AI rankings, achieving a score of 71.4 in the Voice Design Benchmark and high marks in accent modeling (60.8). The new features are available for creators and developers in Google AI Studio, as well as via the Gemini API, Gemini Enterprise, Gemini Notebook, and Google Vids.

For AI builders: the new models significantly expand capabilities in personalization and voice expression control, offering script editing tools that allow linear control over audio parameters in long-form audio.

OpenAI introduces MentalHealthBench

According to an announcement on September 23, OpenAI presented MentalHealthBench, an expert-driven benchmark. This tool is designed to evaluate helpful and safe responses generated by artificial intelligence within realistic mental health conversations.

For AI builders: The new benchmark allows for a systematic way to verify the quality and safety of conversational system interactions in this sensitive thematic area.

Ringg deploys GPT-5.6-based AI agents

Ringg announced on September 23 the deployment of multilingual AI agents that take over customer call handling. The system uses the GPT-5.6 model to manage communication via voice, chat, WhatsApp, and websites. This solution resolves up to 65% of incoming customer requests. According to the company, utilizing GPT-5.6 reduced operational costs by 90% compared to the GPT-4.1 model.

For AI builders: this example highlights significant progress in the cost-efficiency of newer model generations in specific customer service business applications.

OpenAI and Grab collaboration in Southeast Asia

OpenAI and Grab announced the launch of a regional program called GO Forward with AI, according to information from September 23. This initiative aims to support 30,000 partners in acquiring practical artificial intelligence skills. The program is targeted at users in Southeast Asia and focuses on transferring the knowledge necessary to effectively use AI solutions in daily work and developing digital competencies in the region.

For AI builders: this partnership demonstrates how technology companies scale education on practical AI tool applications for massive groups of end-users.

Microsoft optimizes inference infrastructure for physical robotics

Researchers from Microsoft Research published analysis results on September 23 showing that offloading artificial intelligence inference processes from a robot's onboard computing units to external infrastructure (edge or cloud) significantly increases machine work efficiency.

According to the study, the previous model relying solely on built-in graphics processing units (GPUs) limits performance, battery life, and the scalability of robotic systems. Tests conducted on mobile manipulation tasks showed that offloading computations enables the handling of more complex AI models while simultaneously shortening reaction times and increasing action accuracy. In tests utilizing lighter onboard hardware supported by external computing power, it was observed that relying solely on built-in, less powerful GPUs could lead to a 50% drop in model accuracy and planning process slowdowns of up to 383% compared to an A100 unit.

An additional benefit is a significant extension of robot operating time without recharging. Using external systems allows for the replacement of energy-hungry onboard components (such as Jetson Thor, whose use can drain the battery 160% harder) with lighter solutions like Raspberry Pi-5.

As part of expanding the Physical AI Toolchain, Microsoft introduced new tools supporting this working model. Developers can use Kubernetes-based software enabling containerization and automatic workload distribution between the robot, edge infrastructure, and the cloud. This framework is fully integrated with robotics simulators, LeRobot, and ROS2, and the published documentation provides sample projects for SO-101, UR10e robots, and the Rho model controlling the Mobile Aloha robot.

For AI builders: Deploying robotic systems requires balancing computational performance with energy constraints and network latencies, which is why offloaded inference solutions are becoming a key element in designing infrastructure for advanced robotics operating in unpredictable conditions.

Frequently asked questions (FAQ)

What capabilities does Gemini 3.8 Flash TTS offer?

Gemini 3.8 Flash TTS enables the creation of new voices using prompts and precise control over performance parameters such as pace, emotion, and speaking style. The model supports over 100 languages and dialects, allows for the replication of existing voice timbres, and supports multiple speakers in a single audio scene.

Why is offloading computations from robots to external infrastructure beneficial?

Offloading AI inference processes from onboard GPUs to external infrastructure increases performance, action accuracy, and battery life. This allows for the use of lighter onboard hardware, prevents drops in model accuracy, and shortens the reaction time of robotic systems compared to operating solely on built-in units.

Sources

• Gemini 3.8 text-to-speech says hello — 2026-09-23
• Introducing MentalHealthBench — 2026-09-23
• Ringg’s AI agents resolve up to 65% of customer calls with OpenAI — 2026-09-23
• Grab and OpenAI bring practical AI skills to Southeast Asia — 2026-09-23
• Offloaded inference for real-world physical AI robotics — 2026-09-23

--- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.