Google · AI / ML · Filed Jul 13, 2026 · USPTO published Nov 12, 2026 · This week's release

Google Patents a Way to Teach Speech AI New Languages Without Retraining It

Instead of retraining a speech AI for every new language, Google describes leaving it untouched and bending the incoming sound until a Lithuanian 'ne' looks like an English 'no.'

By Patentlyze Team · Updated Nov 12, 2026 · 3 min read

A phone turns spoken questions into text on screen and can send them to a remote server. Drawing from patent filing US 2026/0360354 A1. — A phone turns spoken questions into text on screen and can send them to a remote server.All 7 drawings ↓

Publication number: US 2026/0360354 A1(US20260360354A1)
Applicant: Google LLC
Filed · Published: Jul 13, 2026 · Nov 12, 2026
Inventors: Bo Li, Yu Zhang, Mr. Nanxin Chen, Rohit Prabhavalkar, Mr. Chao-Han Huck Yang, Tara N. Sainath, Trevor Strohman
Status when we published: Waiting for an examiner
Document: 20 claims
US classification: 704/232
Parent application: Continuation of 18/490,808 (filed Oct 20, 2023)(18490808)

What Google's language-adapter patent actually does

Today, a voice assistant that has mastered English can't automatically follow Lithuanian. Teaching it a new language usually means a long round of extra training, and the retrained model can end up too big to run on a phone. Languages with little recorded speech to learn from get left behind.

Google's filing takes a different route. It leaves the English-trained model completely untouched and bolts on two small add-ons that are the only parts that learn. One reshapes the incoming audio so a Lithuanian 'ne' sounds more like an English 'no.' The other nudges the model's internal signals so the two words look alike once they're inside.

Because the original model never changes, it keeps what it already knows. And the new pieces are small, so the whole package stays light enough for a phone.

From the filing · CLAIM 1

obtaining a pre-trained automatic speech recognition (ASR) model configured to generate a latent projection output and a corresponding speech recognition result from an input audio sequence, wherein parameters of the pre-trained ASR model are frozen …

Translation: Google starts with an existing speech recognition model and locks its core settings so they cannot change.

How the frozen model gets two small add-ons

The starting point is a speech recognition model (software that turns spoken audio into text) that is already trained, in the filing's example on English. Its internal settings, called parameters, are frozen, meaning training never changes them. Google then attaches two small trainable pieces.

Input module: adds a learned 'continuous embedding' (a set of numbers tuned during training, which the description also calls trainable noise) to the incoming audio, nudging it toward speech the model already knows.
Latent module: trainable layers inserted into the encoder (the part that turns sound into the model's internal representation) that adjust those internal signals so a Lithuanian word lands near its English counterpart.

Each add-on mixes two ingredients: one that depends on the specific sample in front of it, and one universal piece that stays the same for every sample. The description gives each new language its own universal piece. In training, the model transcribes a sample, the result is scored against the correct transcript (the loss), and backpropagation (the standard method for working out which numbers to adjust) updates only the two add-ons.

The filing also describes optional extras: bridged connections between the added layers, and learning from untranscribed audio by hiding chunks of the signal and checking that the model reaches the same answer.

From the filing · THE ABSTRACT

… adapting the ASR model to learn how to recognize speech in the second language by training the input reprogramming module and the latent reprogramming module while parameters of the ASR model are frozen …

Translation: The system teaches the AI a new language by training only small add on modules instead of retraining the whole model.

Why smaller language add-ons matter for phones

If you speak a language with little recorded speech online, voice typing and assistants tend to work worse for you. This filing describes a cheaper way to close that gap: instead of building a whole new model for each language, you add a small package of learned pieces on top of one that already exists. The filing's own argument is that this keeps the model small enough to run on a device, so recognition can happen on the phone itself.

It also suggests one base model could carry add-ons for several languages, since the description covers training on speech from other languages too. The filing names no product and gives no accuracy numbers in the text we reviewed, so how well it performs is an open question.

Google's 67th filing in the voice and speech AI work we've tracked since May adds to a run that includes one on noise-free voice commands and one on fixing misheard names.

Editorial take

Claim 1 is wider than the title suggests. It talks about a 'target domain' that differs from the model's 'source domain,' and language only shows up in the later claims. So the core idea reaches any new kind of audio the original model was not trained on, not only new languages.

The build it requires is very specific. The model must stay frozen, one add-on must sit at the input, trainable layers must sit inside the encoder, and training may change only those add-ons. The inner add-on must also combine data tied to each sample with data that stays the same across samples. A system that skips one of those pieces would fall outside the wording as written.

Overall it is a practical filing about making speech software cheaper to extend. The claim covers many kinds of new audio but demands one particular recipe to get there.

Source: US 2026/0360354 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.

Add Patentlyze on Google One click.

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 · 7 sheets

7 drawing sheets from US 2026/0360354 A1 · click any drawing to enlarge

How a phone or server transcribes your spoken request, drawing sheet 1 from US 2026/0360354 A1 — Sheet 1 How a phone or server transcribes your spoken request

The speech model converting spoken audio into words, drawing sheet 2 from US 2026/0360354 A1 — Sheet 2 The speech model converting spoken audio into words

Training the speech model on labeled voice clips, drawing sheet 3 from US 2026/0360354 A1 — Sheet 3 Training the speech model on labeled voice clips

Tuning the same model using uncaptioned voice recordings, drawing sheet 4 from US 2026/0360354 A1 — Sheet 4 Tuning the same model using uncaptioned voice recordings

Connecting helper parts to update the voice system, drawing sheet 5 from US 2026/0360354 A1 — Sheet 5 Connecting helper parts to update the voice system

Steps for adapting voice recognition to a new language, drawing sheet 6 from US 2026/0360354 A1 — Sheet 6 Steps for adapting voice recognition to a new language

Computer hardware that powers the voice recognition system, drawing sheet 7 from US 2026/0360354 A1 — Sheet 7 Computer hardware that powers the voice recognition system

The drawings · 7 sheets

More from Google

Google Patents a Speech Recognition System That Understands Multiple Languages Simultaneously

Google Patents a Speech Recognition Model That Trains Across Languages Using One Alphabet

Google Files Patent to Cut the Computing Cost of AI-Generated Speech and Audio

Google Files Patent for Teaching Speech Recognition to Learn From Its Own Best Guesses

More in AI / ML

Nvidia Patents a Two-Stage Speech Transcription System That Handles Multiple Languages at Once

Nvidia Patents a Way to Train One Speech AI Across Dozens of Languages at Once

Nvidia Files Patent to Stop AI From Forgetting Text While Learning Speech

A New Microsoft Patent Trains Speech Recognition on Voices Scrubbed of Identity

No comments yet

Start the discussion Cancel reply