Nvidia Patents a Single AI Model That Transcribes and Translates Speech at Once
Most speech tools make you pick between transcription and translation as separate steps. Nvidia's new patent covers a single model that does both at the same time, turning spoken words directly into text in whatever language you ask for.
What Nvidia's combined speech-and-translation model actually does
Today, if you want to transcribe audio and then translate it into another language, you typically run two separate tools, one after the other. That chain adds delay, cost, and more places for errors to creep in. Nvidia's patent describes a way to collapse both jobs into a single AI model.
You speak into the system, tell it which language you want the output in, and the model produces written text in that target language directly. No handoff between systems, no intermediate transcript that gets passed to a separate translator.
The key idea is that the model learns a kind of blended understanding of your speech and the target language at the same time, combining those two signals before it writes out the final text. The result is a single pipeline that handles multilingual speech recognition and translation as one unified task.
… generating one or more combined vectors based at least on the one or more audio vectors and the one or more language vectors …
Translation: The system merges the spoken audio data with the chosen target language data.
How the model merges audio and language signals into one output
The patent describes an end-to-end multilingual speech model built around three stages.
- Audio encoding: One or more encoders take raw audio data and convert it into a set of numeric vectors (think of vectors as a compact mathematical fingerprint of the sounds and words in the audio) inside a shared embedding space.
- Language encoding: Separately, the model generates vectors representing the target language the user wants the output in. These language vectors capture the structure and vocabulary of the destination language.
- Combining and decoding: The model merges the audio vectors and the language vectors into combined representations, using operations like concatenation (stacking them together), masking, or weighted addition. A decoder then reads those combined representations and outputs text in the target language.
Because the audio and language signals are fused before the decoder runs, the model doesn't first produce a transcript in the source language and then translate it. It reasons about both dimensions simultaneously, which is where the efficiency claim comes from.
The claim language is broad: it covers any machine learning model that follows this encode-combine-decode structure, without specifying a particular architecture, size, or number of languages.
… combine the audio representation(s) with the language representation(s)-such as by performing concatenation, masking, fusion, adding, and/or the like-to generate one or more combined representations.
Translation: The model blends the speech and language data together using mathematical techniques.
What this means for real-time translation in Nvidia-powered products
For Nvidia, whose hardware already powers a large share of AI inference workloads in data centers and edge devices, a patented approach to single-model speech-plus-translation could matter for products ranging from real-time meeting transcription tools to in-vehicle assistants. Nvidia's push into AI-powered audio and speech has been building steadily, and owning IP around the core pipeline architecture gives the company a potential licensing or product-differentiation lever.
For you as an end user, the practical upside is speed and accuracy. Removing the translation step means fewer words get garbled in handoff, and the system can respond faster, which matters most in live conversation or voice-driven interfaces where every second of lag is noticeable.
Nvidia's 18th filing in the voice and speech AI patents we've tracked since May adds to a run that includes separating speakers in group calls and identifying who spoke by content.
Claim 1 covers a method that takes recorded speech, converts it into a numerical summary, generates a separate numerical summary for a target language, combines the two, and outputs translated text. Written at that level of generality, the claim describes the basic sequence of steps that underlies virtually every voice translation tool in use today.
If granted at this scope, Nvidia would hold legal rights over a broad category of software behavior rather than a specific technical invention. Any application that encodes speech, applies language conditioning, and generates translated text could fall within the claim's reach.
Patents written this broadly typically face pressure during government review to narrow down to specifics, such as the precise method used to combine the speech and language summaries. How much of this claim survives that process will determine whether the patent carries real market weight.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
14 drawing sheets from US 2026/0260655 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →