NVIDIA unveils Voice Foundry, a dedicated suite of NIM inference microservices for TTS and STT optimized for Blackwell GB200 hardware, promising sub-80ms first-token latency and 70% lower per-character costs for enterprise voice applications.
At its GTC 2026 Spring event, NVIDIA announced Voice Foundry, a new family of NIM (NVIDIA Inference Microservices) dedicated exclusively to speech AI workloads. The service packages pre-optimized TTS, STT, and voice-cloning models — including partner models from ElevenLabs, Cartesia, and Meta's newly released Llama-Voice — into drop-in containers tuned for the Blackwell GB200 NVL72 platform. NVIDIA reports first-token latency as low as 78 milliseconds and a 3.1x throughput improvement over the same models running on H100 Hopper GPUs.
The economic case is the headline. NVIDIA claims that enterprises running Voice Foundry on GB200 reference systems can serve real-time voice generation at approximately $4.50 per million characters, roughly 70% below the list price of comparable public cloud TTS APIs. The savings come from a combination of Blackwell's FP4 inference support, TensorRT-LLM speculative decoding optimized for audio tokens, and a new 'audio KV cache compression' technique that reduces memory bandwidth requirements by 4x during streaming generation. Early adopters Revolut and Zoom confirmed migrations that delivered 65–78% cost reductions in internal pilots.
Voice Foundry is also a pointed competitive move against hyperscaler voice APIs. Rather than competing directly with OpenAI or Microsoft on model quality, NVIDIA is positioning itself as the neutral infrastructure layer: any voice model, any enterprise, any cloud or on-premise deployment. The launch includes official partnerships with ElevenLabs (optimized deployment of their v3 multilingual model), Cartesia (Sonic-2 streaming), and Resemble AI (voice cloning with enterprise consent workflows). Each partner model is offered with validated reference architectures and commercial SLAs.
Industry observers see Voice Foundry as the infrastructure signal that voice AI has crossed from experimental to production-scale workload. 'Whenever NVIDIA dedicates a NIM category to something, it means the workload is no longer rounding error in data center capex,' wrote analyst Dylan Patel. The launch coincides with a broader push by NVIDIA to extract more value from voice and multimodal workloads, which the company estimates will account for 25% of AI inference cycles by 2028, up from under 6% in 2025.
