Here's what actually issued. On September 8, 2020, Google LLC was granted US10770064B1, "System and method for speech recognition using deep recurrent neural networks," with Alexander B. Graves named as inventor — a researcher closely associated with sequence modeling and recurrent architectures. The CPC codes pair the speech class G10L 15/16 with neural-network classes G06N 3/0445 (recurrent networks) and G06N 3/08 (learning methods).
The mechanism is sequence modeling. Speech is a time series — a stream of acoustic frames that must be mapped to words — and recurrent networks process sequences by carrying state forward from one step to the next. The claimed system uses deep recurrent networks to perform that acoustic-to-text mapping. This is the architectural generation that powered voice assistants and dictation before transformer-based speech models became standard.
The publication and grant timing is worth noting for the prosecution-minded reader: a B1 grant in 2020 reflects an application filed years earlier, when recurrent networks were the dominant tool for speech. The patent captures that moment. It is a reminder that grants are historical artifacts of the filing date as much as descriptions of current practice.
On scope, the boundary is the claim language. Granted and enforceable, but the independent claim describes a specific recurrent speech-recognition method. It does not foreclose the entire field, and it certainly does not reach the transformer-based approaches that came later. Don't read the broad title as the monopoly.
The takeaway: US10770064B1 is Google staking a claim on a core enabling technology for voice products, naming a marquee sequence-modeling researcher, with claims pinned to the recurrent method that was state of the art when it was filed.