Nvidia unveils Magpie TTS model with open weights and full infrastructure control

The voice-processing pipeline consists of many stages, from audio capture through speech-to-text (ASR), large language model (LLM) inference, and finally text-to-speech (TTS), the last stage being the one users notice instantly. If speech generation is slow, the entire experience feels sluggish. Nvidia’s Magpie TTS is positioned to give developers complete control over this latency budget and to enable fine-grained tuning of every component in a self-hosted stack.

Integrated speech models often expose a single API that accepts audio and returns audio, but that simplicity trades off the ability to adjust each stage, swap in improved models as they are released, enforce local data residency, and understand the source of latency. A cascaded architecture keeps ASR, TTS and LLM as separate modules that cooperate while remaining independently configurable and deployable by the developer. The new model is released with open weights, ready for production with Nvidia NIM, and supports 12 languages.

The current release expands multilingual coverage to include Modern Standard Arabic, Korean and Brazilian Portuguese. It also raises quality across many existing languages using updated training data and model refinements. Magpie TTS Multilingual is a 364-million-parameter model that adds support for English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi and Japanese. Each language provides both male and female synthetic voices through a shared multilingual speaker representation. The version further improves flexibility with extended code-switching support for Hindi and Japanese, leveraging IPA-based phoneme processing and custom pronunciation dictionaries to ease rendering of names, technical terms and mixed-language content.

In conversational AI systems, time-to-first-audio (TTFA) is the most critical metric in the voice pipeline. Because Magpie TTS can be deployed inside an organization’s internal environment, the server-side latency measured is fully under developer control, without round-trip transport to a managed service. According to Nvidia TTS NIM documentation (version 26.07), based on the average of three local-environment experiments, the model achieves a TTFA of 32 milliseconds on a B200 processor. The source does not provide comparative figures for other models or full end-to-end output latency (RTFX) metrics.