Convert media duration and measured real-time factor into expected per-task latency and safe hourly capacity for an AI transcription service. Compare that capacity with a target workload while reserving utilization headroom, so infrastructure plans reflect queueing risk instead of assuming every worker runs at 100 percent.
Task latency (seconds) = audio minutes × 60 × real-time factor + post-processing seconds. Safe capacity = workers × 3,600 ÷ task latency × planned utilization.
The main result is a planning estimate for one transcription workload or its available capacity. It is only as accurate as the pricing, model-speed, token-density, and workload assumptions entered. Use measured production values whenever possible and run high- and low-case scenarios before committing capacity or a customer price.
This planning estimate does not guarantee vendor billing, model performance, transcription accuracy, translation quality, or service-level objectives.
For 30 minutes of audio, 0.25 RTF, 20 seconds overhead, 8 workers, and 75% utilization, latency is 470 seconds and safe capacity is about 46.0 tasks/hour.
Use representative production values or a weighted average from recent tasks. Peak values are better when testing capacity risk.
Not automatically. Increase workload, overhead, tokens, or utilization headroom to model retries and operational variance.
Yes. Enter each option’s measured speed, infrastructure assumptions, and labor or usage prices as separate scenarios.
Headroom helps absorb variable media length, traffic bursts, startup time, retries, and queueing without immediately missing targets.
Update it after pricing, models, hardware, workflow stages, average task size, or concurrency patterns change.
| Variable | Meaning |
|---|---|
| Workload | Media length, source volume, or concurrent demand entered above. |
| Efficiency | Measured model rate, token density, batching gain, or utilization assumption. |
| Overhead | Workflow work outside the primary model calculation. |
| Result | Estimated requirement or capacity before unmodeled operational variance. |