Deploying an automatic speech recognition system

Build an ASR serving path for uploaded recordings and live sessions. Estimate audio storage and measured model throughput, use the correct Whisper large-v3 features, and define how chunks become a coherent transcript.

The previous lesson examined training and evaluation; it did not establish a production-ready checkpoint. This design assumes a selected, evaluated Whisper large-v3 model and adds the operational behavior needed for a service.

Trace one recording through input handling, feature preparation, task configuration, decoding, and transcript assembly. Measure latency and errors at those boundaries.

Begin with a declared workload.

Resource estimation

Assume 100 million daily active users, each submitting ten thirty-second clips per day. Separately assume 100 million stored profiles. These are hypothetical workload inputs, not observed traffic for an existing product.

Storage estimation

Separate standing artifacts and account records from daily retained uploads and transcripts.

Stable storage