Value of automatic routing on sparse workloads

Investigate whether automatic per-frame routing becomes more valuable on sparser speech-enhancement workloads with long pauses and more frequent opportunities to use a cheap bypass.

Background

The study evaluates dynamic depth primarily on the speech-dense VoiceBank–DEMAND dataset and finds that per-frame depth selection provides little additional benefit over compute-matched static models. The authors attribute the limited benefit partly to the scarcity of long pauses and cheap bypass opportunities in the evaluated workload.

The unresolved question is whether the conclusions change for sparser workloads in which long pauses occur more often, potentially allowing automatic routing to reduce computation more substantially while preserving enhancement quality. The paper identifies this as an important next step rather than answering it experimentally.