NVIDIA Releases Open-Weights Magpie TTS Model Supporting Twelve Languages for Low-Latency Voice Agents

NVIDIA has released the open weights for Magpie TTS, a multilingual text-to-speech framework designed to facilitate the construction of low-latency voice agents. The model supports 12 languages, including English, Spanish, German, Chinese, and Japanese, enabling unified multilingual speech generation within a single model. By providing open weights suitable for commercial deployment, NVIDIA allows developers to maintain full control over their deployment infrastructure and data residency. Unlike traditional integrated speech APIs that bundle all speech processing steps into a single external API call, Magpie TTS employs a cascaded architecture. This design gives developers the flexibility to fine-tune individual components, resolve latency bottlenecks, and optimize the total latency budget across integrated LLM, ASR, and TTS pipelines. It addresses common enterprise challenges associated with black-box proprietary APIs, such as data privacy concerns and rigid pipeline configurations. While Magpie TTS is optimized for real-time speech generation, achieving peak inference speed depends heavily on specific hardware setups and overall pipeline configurations. To achieve target response times under tight latency constraints, developers should benchmark and optimize the model using NVIDIA's dedicated optimization and inference tools within their actual deployment environments.