New Google Patents · Filed Mar 19, 2026 · Published Jul 30, 2026 · verified — real USPTO data

Google Patents a Way to Transcribe Long Audio Without Losing the Thread

Most AI transcription tools struggle when audio goes on for more than a minute or two. Google's new patent describes a way to feed long recordings into a speech model in manageable pieces without losing accuracy as the audio stretches on.

Google Patent: Chunk-Wise Attention for Long Audio Transcription — figure from US 2026/0221131 A1
Figure from the official USPTO publication.
See all 12 drawings from this filing ↓
Publication number US 2026/0221131 A1
Applicant Google LLC
Filing date Mar 19, 2026
Publication date Jul 30, 2026
Inventors Mr. Yongqiang Wang, Pedro J. Moreno Mengibar, Wei Han, Yu Zhang, Parisa Haghani
CPC classification 704/8
Grant likelihood Medium
Examiner CENTRAL, DOCKET (Art Unit OPAP)
Status Docketed New Case - Ready for Examination (Apr 22, 2026)
Parent application is a Continuation of 18585168 (filed 2024-02-23)
Document 20 claims

What Google's chunk-based audio transcription actually does

Imagine asking a friend to transcribe a two-hour meeting recording. If you hand them the whole thing at once, they get overwhelmed. But if you give them one page at a time, they stay sharp. Google's patent works the same way for its speech-to-text AI.

The system takes a long audio recording and cuts it into equal-length chunks before sending each piece through a neural network trained to recognize speech. The network processes each chunk while staying aware of what came just before, so it doesn't lose track of a sentence that straddles a boundary.

The underlying model is also trained on multilingual data, both spoken and written, so it can handle a wide range of languages and accents. The goal is accurate transcription of long-form audio, like lectures, interviews, or phone calls, without the quality drop that typically happens when recordings run long.

How the encoder processes fixed audio slices in sequence

The patent describes a chunk-wise attention mechanism applied to an audio encoder built from stacked self-attention layers (a transformer-style architecture that weighs how different parts of an input relate to each other).

Here's the core pipeline:

  • A long audio recording is converted into a sequence of acoustic frames (short snapshots of sound, typically every 10-25 milliseconds).
  • Those frames are split into fixed-length chunks, keeping the total input manageable for the model at any given moment.
  • The encoder runs self-attention only within and across a limited window of chunks, not across the entire recording at once. This keeps memory use from exploding as audio length grows.
  • A decoder then reads the resulting audio encodings and produces a text transcription.

The training side is also notable. The model is pre-trained on three data types: unspoken multilingual text, un-transcribed real speech (no labels), and transcribed real speech. Each contributes a separate loss term (a measure of error the model tries to minimize), which together shape an encoder that generalizes well across languages and audio conditions.

Contrastive learning (a technique where the model learns by distinguishing real audio features from masked or fake ones) is used on the unlabeled speech data, letting the system learn from raw audio it has never seen labeled.

What this means for real-world speech recognition tools

Long-form transcription is one of the harder unsolved problems in speech AI. Most models are optimized for short utterances, like voice commands or a single sentence, and their accuracy degrades noticeably on anything more than a few minutes. If Google can make this chunk-based approach work reliably, it addresses a real gap in products like Google Meet transcription, Recorder on Android, and Google Docs voice typing.

The multilingual training angle is also worth watching. By pre-training on unspoken text and unlabeled speech across languages, the system could extend accurate transcription to languages where labeled training data is scarce. That matters for you if you record meetings, interviews, or classes in any language other than English and want a transcript you can actually rely on.

Editorial take

This is solid, practical AI research rather than a flashy product announcement. Chunk-wise attention is a known approach to the long-sequence problem, but Google's specific combination of multilingual pre-training objectives and the fixed-chunk encoder architecture suggests real engineering work aimed at production-quality transcription. It's the kind of patent that ends up inside a shipping product within a year or two.

The drawings

12 drawing sheets from US 2026/0221131 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.