Verified, then logged.
Sun 22 Nov 2026 · 00:13 KST / 15:13 UTC한국어 | ENSearch /

Breaking02 / IT

AssemblyAI cuts speech recognition compute costs by 70%

AssemblyAI uses queues to retry interrupted jobs and separates servers by region to keep audio data within geographic boundaries.

A wired headset and a workstation displaying unlabeled audio waveforms sit in front of server rows separated by a glass partition in an unbranded data center. AI-generated image

AWS said on 14 Nov that AssemblyAI cut speech recognition compute costs by 70% compared with on-demand pricing by using spare cloud computing capacity called Spot.

According to AWS, Amazon EC2 Spot capacity can be reclaimed with two minutes' notice. AssemblyAI uses on-demand servers for baseline capacity and Spot servers to handle additional demand.

The AWS post describes a retry system built around Amazon Simple Queue Service (SQS), a service that holds jobs awaiting processing. Jobs that cannot finish before a server is reclaimed return to the queue for another server to process.

AWS said AssemblyAI keeps queues and servers separate in each region to limit the movement of audio data. This arrangement supports data residency requirements, which specify where healthcare, finance and legal customers' data must remain.

The savings figure published by AWS uses on-demand pricing as its comparison. AWS also reported that AssemblyAI's compute demand grew nearly fourfold year on year.

AWS did not disclose figures for the additional processing time caused by Spot interruptions.