Real-Time Factor

Real-Time Factor — A processing-speed ratio defined as computation time divided by media duration, where values below one indicate faster-than-real-time processing.

Acoustic and Modelling Boundary

RTF is a processing-speed ratio, not end-to-end user latency. Buffering, queueing, endpointing, and network delay can dominate perceived latency.

Related Speech-Processing Concepts

• Automatic Speech Recognition
• Throughput
• Latency
• Queueing Theory

Mathematical Boundary Between RTF and Capacity

For one media item, RTF = processing_time / media_duration is a direct ratio. For an ideal serial worker with no additional overhead, the coarse speed ratio is 1/RTF; for example, RTF = 0.1 says that one stream consumes roughly one tenth of its audio duration in processing time. It does not guarantee ten safe concurrent streams.

Service capacity under concurrency depends on arrival rate, service-time distribution, batching, CPU/GPU sharing, and tail latency. Queue stability requires offered load to remain below service capacity, while Little's Law relates L = λW for a stable system. RTF is therefore a model/processing-speed indicator, not an admission-control limit. Applied discussion: Queue Stability in ASR Systems.

This distinction is also important when setting autoscaling and admission-control policies.