Recently, Deepdub announced the launch of its latest speech AI model, Lightning2.5. This model is called the fastest and most scalable speech solution the company has ever developed, designed for real-time production-grade speech experiences. Lightning2.5 demonstrates exceptional support capabilities in application scenarios such as multilingual AI agents, call center automation, real-time content localization, and personalized media and gaming.

Lightning2.5 was developed based on Deepdub's self-developed language and speech foundation models, combining human-level expressiveness and high fidelity, supporting multiple languages. The model has expanded its infrastructure for real-time performance, capable of retaining a speaker's identity, tone, and subtle emotions, thereby providing users with a realistic production-grade speech experience.

Ofir Krakowski, CEO and co-founder of Deepdub, said: "Lightning represents a significant advancement in real-time, expressive speech generation technology. Whether driving multilingual AI agents, immersive content platforms, or large-scale voice-first automation, Lightning can bring foundational speech AI to production applications at the speed, fidelity, and quality that customers need."

This new model comes with a redesigned inference engine, offering 2.8 times the throughput and 5 times the concurrency, with a latency as low as 200 milliseconds, half a second faster than industry standards. This makes Lightning highly suitable for real-time voice interactions, dynamic voice narration, and event-driven AI workflows, particularly applicable to multiple industries such as autonomous voice agents, interactive systems, and large-scale voice automation.

As part of the NVIDIA Startup Program, Deepdub has optimized Lightning for various NVIDIA GPUs, such as H100, A100, L40S, and A10G, supporting deployment in both local and cloud environments. By leveraging the NVIDIA TensorRT-LLM acceleration stack, Lightning achieves high throughput, low latency inference at enterprise scale, while maintaining efficient operational costs.

Richard Kerris, Vice President of Media and Entertainment at NVIDIA, stated: "Organizations are increasingly demanding real-time speech AI that combines realistic quality, speed, and scalability. Deepdub has built Lightning2.5 using NVIDIA GPUs and TensorRT-LLM, enabling multilingual speech experiences with ultra-low latency and high concurrency, helping customers deliver next-generation AI agents, immersive content, and enterprise voice applications globally."

Deepdub will showcase Lightning at the 2025 International Broadcasting Conference (IBC), where live speech localization, dynamic content automation, and scalable voice pipelines will be demonstrated. Those interested can visit the AI exhibition area, Hall 14, Booth 14.B53, to experience the charm of Lightning.

Key Points:  

🎤 Lightning2.5 is Deepdub's latest real-time speech AI model, featuring 2.8 times the throughput and 5 times the concurrency.  

🌍 The new model supports multilingual applications, suitable for scenarios such as AI agents and call center automation.  

⚡ Deepdub will demonstrate Lightning at IBC2025, showcasing its powerful features.