0.01 Cent per Second: Developing a Cloud-based Cost-effective Audio Transcription System for an Online Video Learning Platform

Using automatic speech recognition (ASR) to transcribe videos in an online video learning platform can benefit learners in multiple ways. However, existing speech-to-text APIs can be costly to use, especially for long lecture videos commonly found in such platform. In this work, we developed a cloud-based ASR system that is cost-optimized for the workload of online learning platforms. We characterized such workload and applied a combination of techniques from system architecture, including: (1) serverless, (2) preemptible instance, and (3) batching and audio transcription optimization, including: (1) audio segmentation, (2) cost-based segment merging, and (3) locally hosted transcription model. All of which work together to provide a low transcription cost per minute of audio. We experimented and calculated the processing cost, time, and accuracy and showed that our system offers accuracy on par with existing speech-to-text services at a significantly lower cost. We have also integrated this system into an online video learning platform.